New Frontiers in Practical Risk Management

Issue N. 17- 2019 RGO New FrontiersA in Practical Risk Management

1 Argo Magazine

Iason Consulting ltd is the editor and the publisher of Argo magazine. Neither editor is responsible for any consequence directly or indirectly stemming from the use of any kind of adoption of the methods, models, and ideas appearing in the contributions contained in this magazine, nor they assume any responsibility related to the appropriateness and/or truth of numbers, ﬁgures, and statements expressed by authors of those contributions.

Argo magazine Year 2019 - Issue Number 17 Published in December 2019 First published in October 2013

Last published issues are available online: http://www.iasonltd.com/research

Front Cover: Alberto Burri, Bianco, 1952.

Editors: Antonio CASTAGNA (Managing Partner) Luca OLIVO (Managing Director) Executive Editor: Giulia PERFETTI Graphic Designer: Lorena CORNA Scientiﬁc Editorial Board: Gianbattista ARESI Francesco BONFANTI Michele BONOLLO Antonio CASTAGNA Massimo GUARNIERI Antonio MENEGON Luca OLIVO Giulia PERFETTI Massimiliano ZANONI

Iason Consulting ltd Registered Address: 120 Baker Street London W1U 6TU United Kingdom Italian Address: Corso Europa, 15 20122 Milano Italy Contact Information: [email protected] www.iasonltd.com

Iason Consulting ltd is registered trademark.

Articles submission guidelines Argo welcomes the submission of articles on topical subjects related to the risk management. The articles can be indicatively, but not exhaustively, related to models and methodologies for market, credit, liquidity risk management, valuation of derivatives, asset management, trading strategies, statistical analysis of market data and technology in the financial industry. All articles should contain references to previous literature. The primary criteria for publishing a paper are its quality and importance to the field of finance, without undue regard to its technical difficulty. Argo is a single blind refereed magazine: articles are sent with author details to the Scientific Committee for peer review. The first editorial decision is rendered at the latest within 60 days after receipt of the submission. The author(s) may be requested to revise the article. The editors decide to reject or accept the submitted article. Submissions should be sent to the technical team ([email protected]). LATEX or Word are the preferred format, but PDFs are accepted if submitted with LATEX code or a Word file of the text. There is no maximum limit, but recommended length is about 4,000 words. If needed, for editing considerations, the technical team may ask the author(s) to cut the article.

Issue n. 17 / 2019 3 New Frontiers in Practical Risk Management

Table of Contents

Editorial p. 6

What’s New in the Industry p. 8

Credit Risk

2020 EU-wide EBA Stress Test: A Objectives and Key Aspects of 2020 Methodological Analysis on the Methodology p. 13 Credit Risk Perspective Regulatory Changes p. 14 Comparison between 2020 and 2018 Methodology p. 18 Critical Points for Future Consideration p. 24 References p. 26

NPL Classiﬁcation a Random Forest About the Author p. 28 Approach Introduction p. 29 Massimiliano Zanoni Tree-Based Clustering p. 30 Dataset Description p. 36 Development Approach p. 40 Results and Conclusions p. 49 References p. 52 Appendix p. 53

FinTech

FinTechs and Challenger Banks: About the Author p. 58 Old Business, Brand New Approach Introduction p. 59 Antonio Menegon FinTech at a Glance p. 59 A New Business Model p. 62 Conclusions p. 67 References p. 69

www.iasonltd.com 4 Argo Magazine

Market Risk

Security Market: an Overview of About the Authors p. 71 Repo and Security Lending Introduction p. 72 Transactions Repo and Security Lending p. 72 Nicola Giancaspro and Francesco Zorzi REPO and SEC Lending during the Crisis p. 78 Structured Repo, Tri-Party Repo and the Total Return Swap p. 80 Conclusions p. 86 References p. 87

Issue n. 17 / 2019 5 New Frontiers in Practical Risk Management

Dear Readers,

Welcome to the N. 17 issue of Argo, the last one for this year.

In anticipation of the EU-wide stress tests that will be launched in January 2020, we open the issue with “2020 EU-wide EBA Stress Test: A Methodological Analysis on the Credit Risk Perspective” an interesting analysis of the new EBA methodological guidelines valid for the 2020 stress test exercise by Milica Antonijevic and Tancredi Mollica. The new methodology covers all relevant risk areas and incorporates feedback received during the discussion with the industry in the summer of 2019. In the paper, the authors focus on the Credit Risk area, highlighting the nine main topics where differences with 2018 guideline emerge: scenario reversion, macro-economic projections, provision calculation for NPEs, securitisation exposures, PIT parameters, provision for sovereigns, LTV and impact on REA and IRB regulatory EL.

Always remaining within the field of Credit Risk, Massimiliano Zanoni and a team of Iason colleagues analyse the classification of NPL exposure. In particular, in “NPL Classification A Random Forest Approach” the authors propose a structured statistical approach to classify NPL assets according to their potential recovery level comparing a Machine Learning technique known as Random Forest to a better-known Logit approach. Moreover, they show that the Random Forest approach seems to be as reliable and performing as the more known Logistic approach, providing a solid overall performance even with a limited set of information.

In the FinTech section, you can ﬁnd our latest overview on the FinTech world and challenger banks: “FinTechs and Challenger Banks: Old Business, Brand New Approach”. The new article by Antonio Menegon and Ilaria Biondo focuses on banks business models and customers. They assess whether these players are prospering more thanks to consolidated areas for traditional banks or through services that generate new revenue streams.

In the last section, devoted to Market Risk, Gianmarco Dalessandro, Nicola Gian- caspro and Francesco Zorzi present an overview of the Repurchase Agreement and the Security Lending markets analyzing how these kinds of instruments behaved and which vulnerabilities appeared during the last ﬁnancial and banking crisis.

www.iasonltd.com 6 Editorial

Finally, we would also like to remind you of our newsletter service – this is a monthly update on the most relevant topics about Risk Management. If you are interested and have not yet received our newsletter suggestion, we invite you to subscribe on our website.

We would also like to take this opportunity to wish you all happy holidays.

Enjoy your reading!

Antonio Castagna Luca Olivo Giulia Perfetti

Issue n. 17 / 2019 7 What’s New in the Industry What’s New in the Industry

2020 EU-wide Stress Test Methodology

The European Banking Authority (EBA) published the final Eu-Wide methodology and draft templates for the 2020 EU-wide stress test Stress along with the key milestones of the exercise. The methodology and Test templates cover all relevant risk areas and incorporate the feedback received during the discussion with the industry in the summer of 2019. The stress test exercise will be formally launched in January 2020 and the results published by 31 July 2020. read more

Source European Bank Authority

Date November 2019

Guidelines on Loan Originator and Monitoring

The EBA developed the Guidelines on loan origination and monitoring in response to the European Council Action Plan on tackling the high level of non-performing exposures. The European Council, in its July 2017 Action Plan, invited the EBA to “issue detailed guidelines on banks’ loan origination, monitoring and internal governance which could in particular address issues such as transparency and borrower affordability assessment”. read more

Source European Bank Authority

Date June 2019

German Bond Yields and Debt Supply: Is There a “Bund Premium”?

Are Bunds special? This paper estimates the “Bund premium” as the difference in convenience yields between other sovereign safe assets and German government bonds adjusted for sovereign credit risk, liquidity and swap market frictions. A higher premium suggests less substitutability of sovereign bonds. We document a rise in the “Bund premium” in the post-crisis period. We show that there is a negative relationship of the premium with the relative supply of German sovereign bonds, which is more pronounced for higher maturities and when risk aversion proxied by bond market volatility is high. Going forward, we expect German government debt supply to remain scarce, with important implications for the ECB’s monetary policy strategy. read more

Source International Monetary Fund

Date November 2019

Issue n. 17 / 2019 9 New Frontiers in Practical Risk Management

ISDA SIMM Methodology

ISDA has published the ISDA SIMM™ Methodology, version 2.2. This version of SIMM includes updates based on the full recalibration and industry backtesting of the methodology. It also includes additional granularity for the FX asset class, the removal of curvature margin from equity volatility indexes and an alteration to allow for annual calibration of credit non-qualifying intra-bucket correlations. read more

Source ISDA

Date September 2019

Basel Committee Publishes Consultation Paper on Revisions to the Credit Valuation Adjustment Risk Framework

Improvements to the capital framework to better capture CVA risk is one of the key elements of the Basel Committee's overall efforts to reform global regulatory standards in response to the global financial crisis. This consultation document proposes a set of targeted adjustments to the credit valuation adjustment (CVA) risk framework issued in December 2017. read more

Source Bank for International Settlements

Date November 2019

Capital Requirements for Banks Are Levelling Off

Andrea Enria, Chair of the ECB’s Supervisory Board, says banks have become more resilient and supervisory expectations are stabilising. He also discusses mergers and why we need more transparency in supervision. read more

Source European Central Bank

Date November 2019

www.iasonltd.com 10 Essential Services for Financial Institutions Iason is an international ﬁrm that provides Financial solutions on Risk Management. In particular, Iason is a leader in Governance and advanced Methodology, offering a unique blend of know-how and expertise in the measurement and the management of Market, Credit and Liquidity risks. Furthermore, Iason provides a suite of essential software solutions to meet the Technology needs of its Clients.

Governance Expertise in the activity of Integrated Governance Processes (IGP) in the Risk Management ﬁelds: this holistic approach is called Business-to-Risk Integration (B2R-i) process and is characterized by

soundness in order to ensure the appropriate and consistent methodologies across risks

effectiveness so as to ensure a proper use of the integrated platform in the decision- making and management processes at all levels of the bank comprehensiveness to ensure a really integrated governance that considers all the relevant aspects: data, models, IT.

Methodology We design, estimate, implement and validate the models and the metrics used in the Risk calculations for the following core risks:

Credit Liquidity Market IRRBB Risk Risk Risk

Implemented models cover the scenarios generation, the pricing calculation, and the aggregation phases.

Technology To meet business and regulatory needs of financial institutions, we design, develop and implement advanced software solutions for

modelling scenarios calculating metrics integrating risks

Website [email protected] LinkedIn Credit Risk

2020 EU-wide EBA Stress Test: A Methodological Analysis on the Credit Risk Perspective What’s New in the Industry

2020 EU-wide EBA Stress Test: A Methodological Analysis on the Credit Risk Perspective aaaa

Milica Antonijevic ∗ Tancredi Mollica *

his paper addresses the main novelties brought by the 2020 EU-wide Stress Test Methodological Note. These variations are mainly a consequence of results of the last Stress Test exercise as well as the T regulatory changes (new DoD, new securitization framework and CRR II as well as the prudential back-stop rule), which have already or will come into effect during the timespan of the exercise. In comparison to the previous methodology (i.e. 2018 EU-wide Stress Test Methodological Note), the paper analyses the main topics on which clear differences emerge: scenario reversion, macro-economic projections, provision calculation for NPEs, securitisation exposures, PIT parameters, provision for sovereigns, LTV and impact on REA and IRB regulatory EL. In addition to that, the paper provides a constructive overview of the critical points for the future consideration of the Stress Test methodology.

he first Stress-testing exercise was con- The aim of this paper is to investigate the nov- ducted in 2009 in response to the 2008 elties brought up by the 2020 EU-wide Stress T Financial crisis. The aim was to increase Test Methodological Note [3]. Therefore, in the the availability of aggregate information on the following Sections we will deep-dive into the capacity to recover among banks in the euro- main differences between the 2018 and 2020 zone, as well as to foster the convergence to- methodology regarding the credit risk part. To wards more prudential practices among Euro- do so, the Section “Objectives and Key Aspects pean banks. So far as much as 7 Stress Tests of 2020 Methodology” discusses the main ob- have been conducted. In December 2018 the jectives and aspects of the 2020 Methodologi- European Banking Authority (hereinafter EBA) cal Note [3] in general. Furthermore, Section announced that the next Stress Test exercise will “Regulatory Changes” discusses the main reg- be carried out in 2020. This announcement was ulatory changes that had an influence on the preceded by publication of 2018 EU-wide Stress part of the Note dedicated to credit risk and Test results which confirmed that the banks’ ef- finally, Section “Comparison between 2020 and forts in building a more resilient capital base 2018 Methodology” makes a detailed compar- have actually born fruits(see results reported in ison between the 2018 [10] and the 2020 Stress the Table1). Although the depletion of capital test credit risk methodologies [3]. was higher than in previous two exercises because of the severity of the scenarios and IFRS 9 implementation, the other capital indicators Objectives and Key Aspects of were improved. For instance, the capital build- 2020 Methodology ups are expected to lead to higher average capital levels by 2020 (134bps higher average CET1 As previously mentioned, the pursuit of EU- than in 2016). Also, a 405bps capital impact wide stress test exercise is to create for all mar- across all EBA banks was estimated relative to ket participants- banks, supervisors and others- 388bps in 2016. Still, the NPAs and the low prof- a common analytical framework within which itability remain a key concern. it will be possible to assess and compare the ∗At the time of the writing of this article, the authors were working for Iason Consulting.

Issue n. 17 / 2019 13 New Frontiers in Practical Risk Management

Year Avg. CET1%,Fully loaded starting-point Avg. CET1% Fully loaded, Adv. Scenario Avg. CET1 depletion, Adv. Scenario 2018 15.90% (2017 restated) 11.85% (2020) 405bps (2017 vs. 2020) 2016 14.39% (2015) 10.51% (2018) 388bps (2015 vs. 2018) 2014 11.28% (2013) 8.70% (2016) 258bps (2013 vs. 2016)

TABLE 1: EBA ST results over the last three exercises; source: EBA

ability of EU banks and the banking system are applied spanning from 2020 to 2022. The to resiliently respond to usual and adverse dis- adverse scenario is developed by the ESRB and tresses and capital-position challenges. the ECB in cooperation with EBA and national This is achieved by using the system of tem- banks, whereas the baseline macroeconomic sce- plates which allows to rigorously assess the nario is provided by ECB. The use of new in- banks positions by comparing the starting point ternal models and modification of existing ones data and the results stressed via internally con- is mandatory if these have been approved by sistent methodology and scenarios. In partic- the competent authority by 31 December 2019. ular, this article is dedicated to examining the Also, the Stress test should be run according methodology used within this common frame- to the new definition of default if the bank has work and to discuss the most important aspects implemented it by 31 December 2019. of this methodology. The 2020 stress exercise will be conducted under the assumption of a static balance sheet as in previous exercises, and the projections Key Aspects over the scenario period are to be carried out Just like previous years, the Stress Test exercise based on the accounting regime applicable on is to be conducted in a bottom-up fashion2. The 31 December 2019. Therefore, the changes in sample of banks that participate in the EU-wide the accounting or tax regimes that take place after the launch of the exercise are not to be stress test covers roughly 70% of the euro-area 5 banking sector, all non-euro EU Member States taken into account . and Norway. To be included, the banks need to have at least EUR 30 bn in assets3. For the exercise the highest level of consolidation is required Regulatory Changes and its perimeter is defined by the CRR/CRD. If the scope of consolidation or the structure of The methodology was already significantly the bank are affected by a certain major event modified with the 2018 [10] exercise (as regards previous to the start of the exercise, the bank is the credit risk) with the aim of factoring in the allowed to use the pro-forma data4. The events requirements introduced by International Finan- characterized as “major” are those affecting the cial Reporting Standard 9 (hereinafter IFRS 9). representativeness of the bank’s financial state- The changes within the 2020 methodology, on ments (i.e. more than 12.5% of Total Assets) the other hand, have the goal of addressing the significantly, with the focus on the following: regulatory changes that will enter into force during the time horizon of the stress test (2020- • Mergers; 2022) or before. The new regulatory require- • Acquisitions; ments concern the application of the CRR/CRD amendments that were taken before 1 January • Spin-Off of Relevant Business Units; 2020, which are listed and discussed in more detail in the following three Subsections (Subsec- • Divestments. tions “New Definition of Default”, “New Secu- For the 2020 exercise the figures of 2019 (year- ritisation Framework and CRR II”, “Prudential end) are used as the basis and the scenarios Backstop”). 2This means that the exercise is performed by banks with supervision of the competent authorities. On the contrary, in the US the stress-testing is conducted in a top-down manner by the FED based mostly on aggregate institution data and less detailed information. 3According to the 2020 Methodological note draft, the competent authorities have a discretion “to request to include additional institutions in their jurisdiction provided that they have a minimum of EUR 100 bn in assets.” 4The pro-forma data are the adjusted historical data for the year in which the major event took place or more, which is supposed to reflect the major event in question within the caps and floors prescribed in the methodological note. The competent authorities should send to EBA the list of major events before the first submission, this way the banks entering into the scope will be able to use the pro-forma data in their submissions. 5Unless “they are known to be legally binding during the stress test time horizon and if the requirements (including their implementation schedule) have been endorsed and publicly announced by the relevant authority.” [3]

www.iasonltd.com 14 What’s New in the Industry

New Definition of Default and 500 for non-retail ones. The European Banking Authority (EBA) has fi- – The relative threshold should be set nalized the default definition under Art. 178 at the level of 1% for both retail and CRR [6] and the requirements which enter into non-retail exposures. However, if a force on 1 January 2021 will apply both to competent authority considers that banks using the internal-ratings-based approach this suggested level of the materiality (IRBA) and those using the common reporting threshold does not reflect a reason- standard approach (CRSA). The new Defini- able level of risk it may set a relative tion of Default (DoD) aims at harmonising the threshold at a different level, which approaches used among European institutions in any case must be lower than or to detect defaults. The application of the new equal to 2.5%. DoD would probably lead to significant changes As regards to the criteria for deciding on the in default rates and capital requirements. Of unlikeliness to pay of the obligor, the following course this depends on the level of diversity be- are taken as indicators [6]: tween the old rules applied by each institution against the new ones prescribed by EBA. • The credit obligation is assigned a non- According to the new rules a default shall accrued status. be considered to have occurred with regard to a particular obligor when either or both of the • A specific credit adjustment resulting following have taken place [6]: from a significant perceived decline in credit quality subsequent to the initiation • If the obligor is unlikely to pay its credit of the exposure is recognized by the insti- obligations to the institution or any of its tution. associated institutions (a parent company or a subsidiary) in full and without re- • The credit obligation is sold at a material course to realising security. Such expo- credit related economic loss. sures should be classified as “defaulted” • until a factual and permanent improve- The institution consents to a distressed re- ment in credit quality is not observed by structuring of the credit obligation where the institution. The main criteria to define this is likely to result in a diminished fi- an unlikeliness to pay of the obligor are nancial obligation caused by the material defined below in this Section. forgiveness, or postponement, of principal, interest or, where relevant fees. This • If the exposure is past due for more than includes, in the case of equity exposures 90 days on any material credit obligation assessed under a PD/LGD Approach, dis- to the institution, the parent undertaking tressed restructuring of the equity itself. or any of its subsidiaries. The 90-day dead- • line may be prolonged to up to 180 days The institution has filed for the obligor’s in case of exposures secured by residential bankruptcy or a similar order in respect property or SME commercial immovable of an obligor’s credit obligation to the in- property in the retail exposure class, as stitution, the parent undertaking or any of well as in case of public sector entity ex- its subsidiaries. posures. • The obligor has sought or has been placed The materiality threshold set by the com- in bankruptcy or similar protection where petent authority in accordance with point this would avoid or delay repayment of a (d) of Article 178(2) of Regulation (EU) credit obligation to the institution, the par- No 575/2013 is used in defining the ma- ent undertaking or any of its subsidiaries. teriality of the past due obligation. This threshold is a reflection of the risk that Moreover, the new EBA guidelines on de- a competent authority deems reasonable faults establishes a probation period of three [5]: months as a criteria for a return to non-default status. During this period, the effective likeli- – The absolute threshold shall not ex- ness to repay and reliability of the obligor is ceed 100 EUR for retail exposures supervised by checking that the obligor respects 6UTP stands for Unlikely To Pay 7CR-SA banks are those banks using the Standardised Approach

Issue n. 17 / 2019 15 New Frontiers in Practical Risk Management

its payment duties without delays and does not • Linkages between the new securitisation fall into any UTP trigger6. regulation and CRR: capital requirements The New Definition of Default will enter for securitisation positions are now in the into force on January 1st 2021 and apply both CRR proposal, whereas STS (simple, trans- to IRBA and CR-SA 7 banks. This implies that parent and standardized) eligibility crite- the banks must take the new DoD into account ria and other cross-sectoral provisions are when modelling PD and LGD for the stress test listed in the securitisation proposal. Se- exercise of 2020. This is due to the fact that curitisations must respect some binding the scenarios span from 2020-2022. A different requirements to be identified as STS, in detection of defaults requests a replication of order to be characterized by a more tradi- PD and LGD model estimation due to adjust- tional and less risky structure. ments of historical data, which have an impact on risk weights and capital requirements com- • A new hierarchy of approaches: for cal- putation, especially for banks with IRB-models. culating minimum capital needs for se- Since new DoD is more restrictive we expect to curitisation positions the ’Securitisation see higher default rates, i.e. the non perform- Internal Ratings-Based Approach (SEC- ing exposures may increase with more provi- IRBA)’ is at the top of the revised hier- sions required and an erosion of CET1 ratio8.A archy of credit risk calculation approaches 9 backward replication of default rate time series (SEC-IRBA → SEC-ERBA → SEC-SA) . To would involve a recalibration of PD that affects those institutions to which none of these RWA calculations. The evidence of an increase approaches is available (SEC-IRBA, SEC- of default rates depends on the bias between ERBA, or SEC-SA) for a given securitisa- current and new rules on default detection. Fo- tion exposure will have to assign a risk cusing on 2020 EBA Stress Test purposes, the weight of 1250%. A risk weight floor of new DoD may not be already applied in most 15% is set for all securitisation exposures bank for internal modelling, leading to a post- and for all three approaches. ponement of the effect of this regulatory change to the next stress test exercise. • Preferential treatment of STS securitisations: a more risk-sensitive prudential treatment is provided for STS securiti- New Securitisation Framework and CRR II sations. The three approaches are re- calibrated for all tranches in order to gen- A new Securitisation framework has entered erate lower capital charges for positions in into force at the beginning of 2019 and is based transactions qualifying as STS securitisa- on two european Regulations: 2017/2401 [4] tions. This has particular impact on Small and 2017/2402 [13]. Both regulations reorder and Medium-sized Enterprises (SMEs): the chaotic regimentation of securitisation with the aim to harmonise and integrate different – With the aim of facilitating the ac- legal sources. Regulator also tried to organize cess to credit for SMEs, the SME fac- more clearly the opaque securitisation market tor over the overall RWA factor has from a risk management point of view. These been reduced. Firstly, the scope of regulations come along with CRR II, which has application has been extended to ex- been approved by European Parliament in April positions up to EUR 2.5m (instead of 2019 and enhances the CRR on many different EUR 1.5m). Secondly, the coefficient topics, including credit facilitations for small en- has been reduced from 1 to 0.85. terprises through more convenient risk-weight on SME exposures. However, CRR II will not • Caps: enter into force until 2021 and will most probably not interfere with 2020 EBA Stress test. – The maximum risk weight for senior The main methodological changes on securita- securitisation positions: the “look- sions included in regulations 2017/2401 and through” approach allowed only for 2017/2402 are outlined as follows[1]: senior securitisation positions10; 8CET1 ratio measures how much equity (or in simple words, common stock) the bank holds in comparison to its Risk Weighted Assets. Common Equity Tier 1 Ratio = Common Equity Tier 1 Capital / Risk-Weighted Assets 9ERBA is External Ratings-based Approach, whereas SA is Standardized Approach 10The look through approach allows to assign the risk weight to a senior securitisation position corresponding to a weight equal to the exposure-weighted-average risk weight applicable to the underlying exposures.

www.iasonltd.com 16 What’s New in the Industry

– Maximum capital requirements: ex- the minimum coverage calculated using the ta- tend the application of the overall cap ble 2 the following items: in terms of maximum risk-weighted • exposure amounts to originator and Specific credit risk adjustments; sponsor institutions using SEC-ERBA • Additional value adjustments; and SEC-SA11. • Other own funds reductions; • Elimination of special treatment for certain exposures: second-loss or better posi- • For institutions calculating RWAs using tions in ABCP programmes, treatment of the Internal Ratings Based Approach unrated liquidity facilities, additional own (IRBA), the absolute value of the nega- funds securitisations of revolving expo- tive amounts resulting from the calcula- sures with early amortisation provisions. tion of expected loss amounts deducted from Capital Equity Tier 1, which are • Treatment of specific exposures: computed by multiplying the amounts de- – Re-securitisations: a significantly ducted by the contribution of the expected higher risk-weight floor (100%); loss amount for the non-performing exposure to total expected loss amounts for – Senior positions in SME securitisa- defaulted or non-defaulted exposures. tions. • Where a non-performing exposure is pur- These amendments to previous regulations chased at a price lower than the amount are incorporated in 2020 EBA Stress Test, which owed by the debtor, the difference be- provides a specific set of templates on this topic tween the purchase price and the amount and is also interested in quantifying the impact owed by the debtor. of this new framework on the resilience of EU banking system. • Write-offs by the institution registered since the exposure was classified as non- Prudential Backstop performing. The Regulation (EU) 2019/630 [9] published An exception to the above mentioned rules in April 2019, amended the CRR (Regulation are those non-performing exposures that are 575/2013) with regard to the treatment of NPEs guaranteed or insured by an official export (non performing exposures) by introducing the credit agency. In this particular case, a factor of minimum coverage requirement for newly orig- 0 is to be applied until the 8th year of having a inated (i.e. originated after 26 April 2019) loans non-performing status, after which the factor of that become NPEs. Where this requirement 1 is applied. is not met, the difference between the actual In conclusion, from the 2020 EU-wide Stress provisioning and the required one should be Test perspective, which will use the December- deducted from a bank’s own funds (CET1). The 2019 data as a base, the impact of the pruden- aim of this “prudential back-stop” is to address tial backstop rule is expected to be quite low potential under-provisioning, thus preventing a given that it will be applied only on the non- systemic and EU-wide build-up of new NPEs. performing part of the portfolio originated after The minimum coverage requirement in- 26 April 2019. For instance, if we assume, quite creases gradually depending on how long an realistically, that this newly originated portfolio exposure has been classified as non-performing is in-bonis at the beginning of the exercise, then and it varies between “unsecured” and “se- at year 1, the defaulted part will correspond to cured” exposures. Details are given in the Table the 1st year’s default rate of the chosen scenario 2. and so on and so forth for the following years. The level of the applicable amount of insuffi- This defaulted slice will be evaluated in terms cient coverage to be deducted from the Capital of applied vs. required coverage only at the last Equity Tier 1 is determined by subtracting from year of the scenario horizon.

11In other words, the overall cap in terms of maximum risk-weighted exposure amounts should be applied to all originator and sponsor institutions, regardless of the approach they use.

Issue n. 17 / 2019 17 New Frontiers in Practical Risk Management

3rd year 4th year 5th year 6th year 7th year 8th year 9th year 10th year Unsecured 35 100 Secured (by immovable property) 25 35 55 70 80 85 100 Secured (by funded/unfunded credit) 25 35 55 80 100

TABLE 2: Prudential backstop; source: [9]

FIGURE 1: Scenario reversion under 2020 methodology

Comparison between 2020 and new reversion rule (see Figure 1). These risk- parameters are then used in the calculation of 2018 Methodology provisions (see Subsection “Provision Calcula- tion for S2 and S3”). This Section focuses on the comparison between 2018 and 2020 EBA Stress test methodology. The analysis treats the main topics on which a clear Provision Calculation for S2 and S3 difference emerges: scenario reversion, provision calculation for NPEs, securitisation expo- In comparison with the 2018 methodology [10], sures, PIT parameters, provision for sovereigns, the 2020 methodology is focused on the calcula- LTV and impact on REA and IRB regulatory EL. tion of provisions instead of impairment flows. Additionally, when calculating the provisions Baseline and Adverse Scenario Reversion for the non-performing assets (S3) and performing assets with a significant change in the PD While in 2018 the methodology classified in a (S2), the consideration is not given only to the unique bucket S2 and S3, in the 2020 methodol- existing S2 and S3 exposures (as in 2018) but ogy draft, the regulator distinguishes between also to those exposures migrating from the S1 the calculation of ECL for performing and non- to one of the non-performing stages. Therefore, performing, thus, classifying differently S1 and the stock of provisions for the S1 exposures is di- S2 with respect to S3: vided in the provisions for existing (and remain- • S1 and S2: At the end of the scenario hori- ing) S1 exposures and the provisions for new zon, the adverse scenario reverts towards S1 exposures which migrated from S2. By the the baseline scenario with a 6-years linear same token, the provisions for the S2 (S3) are di- reversion. The baseline credit risk param- vided into those calculated on exposures which eters are assumed to stay flat at the end of remained in S2 (S3) and those which migrated the scenario horizon (in 2018 [10] no spec- from S1 to S2 (from S1 or S2 to S3). Finally, ification was given for either of the scenar- the calculation of provisions incorporates the ios for S1, whereas for S2 the methodology forward-looking risk parameters (for instance, was the same also in 2018). for the calculation of the provisions at t+1 we use the LRt+2). This way the provisions in each • S3: Both baseline and adverse credit risk period move in anticipation of risk changes in parameters are now assumed to stay flat future. at the end of the scenario horizon (in 2018 The provisions for S1 and S2 are calculated [10] no specification was given for either for the End of Year (EoY) exposures, therefore, of the scenarios). incorporating the ECL parameters for the next For the sake of example, hypothetical val- year, while in the 2018 methodology [10], the ues have been made up in order to depict the Beginning of Year exposures were used. For

www.iasonltd.com 18 What’s New in the Industry instance, in order to calculate the S2 existing provisions of the year t+1, the following for- ProvS1 − S2Adv(2022EoY) = mula is applied (see [3]): ExpS1 TR1−2 · (2022BoY) Adv(2022) (5) − − = · ( − 2 1 + 5 LT1−2 1 LT1−2 ProvS2 S2(t+1) ExpS2(t) 1 TR(t+1) · LR + · LR 6 Adv(2022) 6 Base(2022) − 2−3 ) · 2−2 TR(t+1) LRLT (t+2) (1) where ProvS2 − S2Adv(2022EoY) are the existing and ProvS1 − S2Adv(2022EoY) are the S2 provisions resulting from new S2 exposures. On the other hand, since the exposures in where ExpS2(t) are the S2 exposures at the the baseline scenario are assumed to remain ﬂat beginning of year t, TR2−1 refers to the 1-year after the scenario horizon,the provisions of the transition probability of S2 exposures to S1, last year (2022) are calculated as follows: TR2−3 refers to the 1-year transition probability of S2 exposures to S3, LR2−2 refers to the LT − = · lifetime ECL parameter for the following year. ProvS2 S2Base(2022EoY) ExpS2(2022BoY) − 2−1 − 2−3 · LT2−2 1 TRBase(2022) TRBase(2022) LRBase(2022) Similarly, the stock of provisions for the new (6) S2 exposures is calculated using the formula

(see [3]): ProvS1 − S2Base(2022EoY) = ExpS1(2022BoY)· − − (7) − 1 2 LT1 2 1 2 TR ( ) · LR ( ) ProvS1 − S2(t+1) = S1 − S2 f low · LRLT (t+2) Base 2022 Base 2022 (2) In order to better examine the difference be- 1−2 tween the old and the new method, the two S1 − S2 f low = ExpS1(t) · TR (3) (t+1) methods were tested on a set of hypothetical S2 parameters and exposures for both the baseline and the adverse scenario. The provisions were calculated according to each of the methods and where ExpS1(t) are the S1 exposures at the the results are illustrated in the Graph 2. The beginning of year t, LR1−2 refers to the life- LT hypothetical provisions are plotted on the y-axis time ECL parameter for the following year and and the time is plotted on the x-axis. From the TR1−2 refers to the 1-year transition probability graph it can be observed that the new provi- of S1 exposures to S2. sion curves (the red and the gray one) shift in anticipation of future “shocks”, while the old As a result of using the EoY exposures ones (dark red and green) move in response and the next year’s risk parameters, the new to those “shocks”. For instance, if we look at methodology results to be much more forward- the provision curve of the 2020-adverse-scenario looking than the one used in the 2018 Stress we see that in the ﬁrst period (2019-2020) it is Test. This also has implications on the pro- less downward-sloping than the 2018-adverse- visions in the last year of the scenario (2022), scenario curve. This is because in the second which now take into account the effects of re- period the risk parameters increase, which is version which starts after the third year of the visible from the movement of the 2018-adverse- scenario (see Subsection “Baseline and Adverse scenario curve. Furthermore, in the second pe- Scenario Reversion”). Namely, for S2 exposures, riod the 2020-adverse-scenario curve is more the provisions of the last year of the adverse upward-sloping than the 2018-adverse-scenario scenario (2022) are calculated as a sum of the one because of the big increase that happens existing and new provisions: in the third period (which is visible from the third period of the 2018 curve). Finally, in the ProvS2 − S2Adv(2022EoY) = last period (2021-2022) the difference between

ExpS2(2022BoY)· the two curves is the most apparent since the 2020-adverse-scenario curve moves down in an- 1 − TR2−1 − TR2−3 · (4) Adv(2022) Adv(2022) ticipation of the reversion that will follow in the 5 1 period 4, whereas the 2018 curve responds to · LRLT2−2 + · LRLT2−2 6 Adv(2022) 6 Base(2022) the period 3 “shock”.

Issue n. 17 / 2019 19 New Frontiers in Practical Risk Management

FIGURE 2: Provision calculation methodology comparison: new (2020) vs old (2018)

For the S3 exposures the provisions are calculated using the same year’s risk parameters as in 2018 [10]. The provisions for the existing S3 ProvSX − S = exposures are calculated based on the first year 3(t+1) (10) · X−3 · X−3 risk parameters. This is given by the perfect ExpSX(t) TR(t+1) LGD(t+1) foresight assumption according to which the loss rate for the stock of existing S3 exposures X−3 where X=1,2 and TR(t+1) refers to the 1-year at the beginning of the exercise stays the same transition rate of SX exposures to S3, LGDX−3 in every year of projection. At each year, the pro- (t+1) visions on existing S3 exposures are calculated refers to the expected loss rate for exposures ExpSX based on the first year’s parameters [3]: that migrate from SX to S3 and finally, (t) are the SX exposures at the beginning of year t. As for the provisions after year 3 of the scenario ProvOldS3(t+1) = horizon, for S3 exposures, both the adverse and [ · 3−3 ] max ExpOldS3(t) LRLT (t+1); ProvOldS3(t) the baseline credit risk parameters assume a flat (8) trend. Therefore, in the last year of the horizon (2022) the provisions for new S3 exposures are where Exp S3 is the S3 exposures at the be- Old (t) calculated in the following way: ginning of the exercise, whereas ProvOldS3(t) is the stock of provisions for existing S3 exposures 3−3 at t and LRLT (t+1) is the ECL parameter esti- ProvSX − S3Scen(2022EoY) = mated for the exposure S3 at the beginning of ExpS1 · TRX−3 · LGDX−3 the exercise, which stays equal throughout the (2022BoY) Scen(2022) Scen(2022) exercise horizon. As for the new S3 exposures, (11) since no release of S3 provisions is allowed for where X is either 1 or 2 and the Scenario is any year or any scenario, the provisions on new either baseline or adverse. S3 exposures are accumulated throughout the stress test horizon. The following formulas depict the methodology to be applied for the calculation of the provisions for new S3 exposures:

ProvNewS3(t+1) = (9) ProvS1 − S3(t+1) + ProvS2 − S3(t+1)

www.iasonltd.com 20 What’s New in the Industry

Securitisation Exposures not fall within the range of any single CQS, the exposure is put in the closest two CQSs in As mentioned in the Subsection “New Securi- terms of RW (for example senior tranche, non tisation Framework and CRR II” the Regula- STS, with RW of 110%, with maturity of 2.5 tions 2017/2401 [4] and 2017/2402 [13] intro- years will be allocated in 0.667·CQS9(5yr) and duced the changes in the securitization frame- 0.333·CQS9(1yr)). The same reasoning is ap- work which as a consequence, were followed plied when the maturity of the exposure lies by the changes in the methodology that will be between 1 and 5 years. Following this, the secu- described in this Subsection. rities are allocated into 3 securitisation buckets The 2020 methodology shifts from the con- for which the REA increase is prescribed. This cept of generic provision to the concept of spe- bucketing is performed in the same way as in cific credit risk adjustments. Namely, the up- 2018 [10], i.e. the allocation is conducted based dated regulation requires to estimate specific on an analysis of the migration volatility of dif- credit risk adjustments (instead of generic pro- ferent products and their origin, distinguishing visions) for securitisation exposures that are not therefore, among three main categories low risk subject to mark-to-market valuation. Therefore, (bucket 1), medium risk (bucket 2), high risk in addition to the exposures subject to Chapter (bucket 3). While in 2018, the banks had to 5 of CRR [6], also exposures subject to specific provide information on the IRB and STA ex- risk adjustments need to be reported in the se- posures per defined risk bucket (actual and re- curitisation template. Generally, under the new stated), this is no longer necessary within the framework all securitisation exposures are to be 2020 framework. The REA reported for each reported net of specific credit risk adjustments. approach and the RW used for the mapping Furthermore, originator positions where no shall consider the impact of the application of significant risk transfer (SRT) has taken place the maximum risk weight and maximum capital are to be treated and reported in the credit risk requirements. templates as in 2018, however, taking into ac- Furthermore, the new methodology gives no count the credit risk mitigation effect in accor- indication on the use of supervisory formula ap- dance with Article 249 of the CRR. At the start- proach (SFA) in the calculation of REA, nor its ing point, i.e. on 31 December 2019, banks are reporting, as was done in 2018. In fact, the cor- required to report exposure values and REA responding templates used for their reporting separately as actual and restated figures. For in the previous exercise, CSV_CR_SEC_IRB_SF the purpose of the stress test it is assumed that and CSV_CR_SEC_OTHER, are not among 2020 the restatement does not affect the SRT achieved Stress Test templates. as of 31 December 2019. Finally, while the 2018 methodology [10] For all regulatory approaches a fixed risk does not specify any particular treatment of weight increase will be applied, whereas in 2018 re-securitisations, the 2020 note specifies that [10], this increase was applied only on “regu- they shall be treated in line with Article 269 of latory approaches based on risk weights”. For Regulation (EU) 2017/2401 [4] and always be this purpose, the restated securitisation expo- reported under the respective credit quality step sures are now mapped into credit quality steps in the securitisation standardized approach part (hereinafter CQSs) from external ratings-based of the template. Also, the impact in terms of look-up tables of Articles 263(3) and 264(3) of REA from the maximum risk weight for senior Regulation (EU) 2017/2401 [4]. This mapping securitisation positions and the maximum cap- is done using the risk-weight, seniority and ma- ital requirement outlined in Articles 267 and turity of the security, as shown in Table 3 and 268 of Regulation (EU) 2017/2401 [4] shall be for STS in Table 4. In case the restated RW does reported as memorandum item.

Issue n. 17 / 2019 21 New Frontiers in Practical Risk Management

CQS Senior tranche (1yr) Senior tranche (5yr) Non-senior tranche (1yr) Non-senior tranche (5yr) 1 15% 20% 15% 70% 2 15% 30% 15% 90% 3 25% 40% 30% 120% 4 30% 45% 40% 140% 5 40% 50% 60% 160% 6 50% 65% 80% 180% 7 60% 70% 120% 210% 8 75% 90% 170% 260% 9 90% 105% 220% 310% 10 120% 140% 330% 420% 11 140% 160% 470% 580% 12 160% 180% 620% 760% 13 200% 225% 750% 860% 14 250% 280% 900% 950% 15 310% 340% 1050% 1050% 16 380% 420% 1130% 1130% 17 460% 505% 1250% 1250% All other 1250% 1250% 1250% 1250%

TABLE 3: Mapping table for risk-weighted exposure amounts under the External Ratings Based Approach (SEC-ERBA); source: [4]

CQS Senior tranche (1yr) Senior tranche (5yr) Non-senior tranche (1yr) Non-senior tranche (5yr) 1 10% 10% 15% 40% 2 10% 15% 15% 55% 3 15% 20% 15% 70% 4 15% 25% 25% 80% 5 20% 30% 35% 95% 6 30% 40% 60% 135% 7 35% 40% 95% 170% 8 45% 55% 150% 225% 9 55% 65% 180% 255% 10 70% 85% 270% 345% 11 120% 135% 405% 500% 12 135% 155% 535% 655% 13 170% 195% 645% 740% 14 225% 250% 810% 855% 15 280% 305% 945% 945% 16 340% 380% 1015% 1015% 17 415% 455% 1250% 1250% All other 1250% 1250% 1250% 1250%

TABLE 4: Mapping table for STS securitisations under the SEC-ERBA; source: [4]

Projected Point in Time Parameters asset class exposure. In case the 10% threshold is exceeded, banks can use a weighted average When it comes to the approach applied for pro- between internal models’ and benchmark’s pa- jection of PIT parameters the new methodologi- rameters, unless the competent authorities pro- cal draft [3] is more rigid on the various points. vide further instructions. By this the regulator The most important differences such as: is reducing the level of discretion of banks and • The introduction of the minimum asset directing them towards a more conservative and class threshold for the use of internal mod- “one-size-fits-all” solution. els; Similarly, for the portfolios for which nei- • The adjustment restriction for the bench- ther the satellite model nor the stressed TRs, mark parameters; LGDs or LRs are available, the banks are, as in the previous exercise (2018 [10]) instructed to • The reversion to the baseline levels of risk use the benchmark parameters provided by the parameters; ECB. It is however, highlighted by the regulator that these parameters shall not be adjusted in • The treatment of the exposures towards any way (i.e. no expert-judgement based ad- the Parent Company; justments or scalings are allowed). Also, it is • The definition of cure rates highlighted that given the bank’s initial choice on whether to apply the benchmark or inter- will be discussed in this Section. nally derived parameters, no diversions will be Firstly, if the bank’s satellite models do not allowed at the later date unless approved by ensure the estimation of all the PD/TR and the competent authorities. Once again, in the LR/LGD parameters for a minimum of 10% 2020 Stress Test exercise the benchmark parame- of the pivot asset class exposure, the benchmark ters will be used for assessment of the potential parameters need to be applied to the entire pivot “over-confidence” of the internal parameter es-

www.iasonltd.com 22 What’s New in the Industry timates. Also, the banks will be subjected to Stress Test methodology indicating a slightly the cross-sectional comparison and asked to re- higher discretion of banks in accounting for the vise their results, shall they prove to be overly sovereign risk. The importance of sovereign optimistic. positions in banks’ balance sheet is still rele- For the projection of the LGD/LR and life- vant and connected with ECB monetary policy time ECL, it is assumed that all future macroeco- and political turmoil across the euro area. A nomic fluctuations are known for the remaining correct risk evaluation of these exposures is cru- lifetime and possible workout periods of the cial to address the systemic risk underlying the exposure (the so-called perfect foresight prin- evident channel between government’s finan- ciple). By the same token, when calculating cial stability and large sovereign exposures of the stage transition probabilities and the corre- European banks. Sovereign still constitutes a sponding loss rates across stages after the stress relevant portion of total assets and their price horizon ends, in the baseline scenario these pa- volatility remains as a potential threat on the rameters are assumed to stay constant, while in market risk side. However, as we have seen the adverse scenario these parameters should from the results of the December 2018 EBA Risk linearly converge to the 2022-baseline level in Assessment Report [9], sovereign exposures are 6-year time for S1 and S2, or stay flat at the 2022- decreasing since June 2016 but persists as a con- adverse level for S3 (see Subsection “Baseline sistent portion of banks’ exposure, standing at and Adverse Scenario Reversion”). EUR 3.0 trillions at June 2018. Regarding the exposures towards a Parent Company which fall under credit risk scope of Loan to Value the Stress Test, according to the 2020 methodological draft [3] the banks are required to be In response to the growing concerns regarding treated at arm’s length. This implies, that for the increasing LTV ratios [8] and their detrimen- the 2020 exercise, the banks will be required to tal impact on LGD and credit losses [1], in the provide the transition and loss rates for these 2020 Stress Test exercise, for the first time so far, exposures as well. the banks are required to report the exposure- Finally, in relation to the estimation of the weighted average of the LTV ratio at loan level cure rates that are necessary for the calcula- for selected real estate exposure classes. tion of the LGD, the new methodological note The ratio will be calculated as the current [3] provides a clear definition of the cure rate exposure (performing or non-performing) di- which was not available in 2018 [10]. According vided by real estate collateral value12 as in the to this definition, the Cure Rate(t) is defined as following formula: the average cure rate observed during a determined period of time for Si exposures reaching S3 within year t. LTV = Exposure/CollateralValue (12)

Provision Calculation for Sovereign Positions For this exercise the performing exposures For what concerns sovereign positions recorded are calculated after substitution effects and after at amortized cost, the ECB provides, for a se- CCF (credit conversion factors), taking COREP lection of countries, a set of stressed TR, LGD (Common reporting framework) as a reference and LR instead of PD and LGD as in 2018 [10]. exposure definition (in 2018 [10] there was no The application of the provided parameters is mention of the reference exposure definition). mandatory for all countries and banks regard- The COREP is the standardized reporting frame- less of whether a country is one of the most work issued by EBA for Capital Directive Re- relevant countries in terms of exposure of the quirements reporting which covers credit risk, bank or not. Zero loss rates are applied only market risk, operational risk, own funds and for exposures to central banks under both base- capital adequacy ratios. When it comes to LTV line and adverse scenarios. For those countries figures, the regulators underline the two main for which the ECB does not provide the risk differences with respect to the reference COREP parameters, on the other hand, the banks are figures and those are given by: required to estimate their own with an adequate degree of conservatism. This specification 1. Different scope under the Stress Test exer- was not mentioned in the previous EU-wide cise; 12The value of the collateral is assumed to be updated according to the particular macroeconomic scenario.

Issue n. 17 / 2019 23 New Frontiers in Practical Risk Management

2. Exposure amounts are aligned with the for S2 and S3”), instead of impairments like in calculation of provisions. 2018 [10]. Also, the expected loss amounts for equity In case of significant difference relative to the exposures need to be reported only in case they COREP figures, the banks are required to justify are deducted in the common reporting frame- them in the explanatory note. work (no such specification in 2018 [10]). Under non performing exposures, the According to the new methodology, compe- methodological note of 2020 ([3]) considers the tent authorities can request from the bank to S3 exposures after the substitution effects and provide, in the explanatory note, the table sum- after CCF (instead of only “S3 exposures” as marizing the exposure values by LTV buckets stated in 2018 [10]). Also here, the COREP is for exposures under standardised approach for taken for the reference exposure definition. SME and non-SME separately. The table should Finally, as regards the denominator of the report the values taking as the reference year LTV ratio, it refers to all funded RE collateral 2019 and for every year of the scenario hori- that is available for covering any of the above zon. For those buckets with LTV larger than mentioned exposures. Moreover, the collateral 100%, the exposures exceeding or being equal of interest is the bank’s share of the CRR/CRD to the market value of the collateral should be eligible collateral. It is not necessary to apply reported. any regulatory haircuts if the collateral value is expected to reflect the evolution of the RE prices in the given macroeconomic scenario. Critical Points for Future

Impact on Risk Exposure Amount (REA) and Consideration IRB Regulatory EL This Section summarizes a few critical points A “no migration rule” is introduced in the 2020 for the future consideration of the Stress-Test document forbidding the migration of expo- methodology. The 2018 stress test has been au- sures and REA between different asset classes, dited by the European Court of Auditors and i.e. the exposure value of each asset class is many recommendations have been provided static meaning that the REA remains in the same [10]. First of all, the impact of the new stress asset class regardless of the scenario migrations. test depends also on the macroeconomic sound- One example of the application of this rule are ness of the adverse scenario. The intensity of the exposures which do not fulfill the “fully and the adverse scenario varies across countries and completely secured” conditions due to wors- for some of them is not stressful enough. This ened collateral amounts. Banks are required difference may harm the consolidation and the to estimate the risk weights (according to CRR) comparison at European level of the final results. and project collateral and credit quality in line Moreover, the critics also point out that the with the scenarios. translation of the adverse scenario into risk pa- Like in 2018 [10], the IRB excesses and short- rameters has not been transparently explained, falls should be accounted for at an aggregate leading to a detrimental lack of information for level for defaulted and non-defaulted portfo- the market participants, which are not able to lios separately. However, the IRB excesses of evaluate correctly if the adverse scenario has credit risk or additional value adjustments over been consistently applied to all banks in a com- expected losses should be considered as Tier parable way. 2 capital, instead of the “excess of provisions” Another issue concerns the focus of EBA on considered in the 2018 Methodological note [10]. economic shocks more than on financial shocks This means that the approach shifts to the calcu- evolving from inner failures of the financial mar- lation of credit-risk adjustments instead of the kets. This approach could lead to an underes- estimation of provisions. The credit risk adjust- timation of the systemic risk of the banking ment development, starting with the reference sector and is emphasized by the exclusion of year and throughout the scenario horizon, is many large banks from the sample. The size now linked to the estimation of provisions on criterion13 excludes many banks on undergoing performing and non-performing exposures (as reconstruction or with critical features in terms described in Subsection “Provision Calculation of drop of consolidated assets, State aid or large 13The overall sample covers the banks that absorbs the 70% of total consolidated assets of euro area banks and with total assets over the EUR 30 billion threshold.

www.iasonltd.com 24 What’s New in the Industry portion of NPL in their balance sheet. namics, a credit growth path conditional on the Moreover, the European and the US stress scenarios could be defined ex-ante (as done in tests should achieve a higher level of conver- UK) or the projections could be adjusted ex-post gence since this would further improve the dis- taking into account the mitigating actions. A cipline in the capital markets. One of the main Static balance sheet lacks reality even if it avoids differences between the two systems lies in the complicating the assumptions and development bottom-up nature of the EU-wide Stress Test of the stress. that allows to take into account the peculiarity Finally, the introduction of very high cov- of single banks, in comparison to the top-down erages for the NPEs presented in the Section US approach that is a more “one-size-fits-all” so- on the prudential back-stop in table 2, might lution, which allows a greater involvement and not have such a strong impact on this Stress discretion of the US authorities. The public has Test exercise because of the limited time-span also expressed its concern regarding the swaps that they apply to, but in the future exercises in which the American and European banks its impact is expected to be fairly high. In par- could engage in the fourth quarter every year. ticular, the requirement to apply the back-stop Namely, since the US and EU stress tests take coverage of 100% on secured exposures could place at different times of the year, the banks be particularly burdensome for some banks. could find a way to temporarily “clean” their In conclusion, the 2020 methodology does balance sheets around times of the Stress test not include relevant changes and revisions from by temporarily swapping the assets. This is also the previous 2018 version, but it only applies a characteristic of the bottom-up nature of the a fine-tuning on the 2018 note and aims to re- Test and of the incentive for a bank to present spect the compliance to most recent regulation the a better picture [14]. updates. This is also due to the strict calendar Another point often criticized is the assump- that exists between the conclusion of the 2018 tion of the Static Balance sheet that reduces the exercise and the drafting of 2020 methodologi- complexity of the exercise but it also introduces cal note. For this reason, the recommendations a lag between the results and the supervisory provided by the European Court of Auditors are decisions. In order to achieve a better trade- supposed to enter in force in the 2022 EU-wide off between the simplicity and more realistic stress test, if embraced by EBA. assumptions regarding the balance sheet dy-

Issue n. 17 / 2019 25 New Frontiers in Practical Risk Management

References [9] European Banking Authority. Risk assessment of the european banking system. December 2018.

[10] European Court of Auditors. Special report No 10/2019: EU-wide [1] Deutsche Bundesbank. Stress stress tests for banks: unparalleled testing the German mortgage market. amount of information on banks 2019. provided but greater coordination and [2] European Banking Authority. focus on risks needed. 2019. 2018 EU-wide stress test Methodological Note. 2018. [11] European Parliament. Regulation (EU) No 2019/630 amending [3] European Banking Authority. Regulation (EU) No 575/2013 as 2020 EU-wide stress test Draft regards minimum loss coverage for Methodological Note. 2019. non-performing exposures. 2019. [4] European Banking Authority. Adoption of the banking package: [12] European Parliament and revised rules on capital requirements European Council. Regulation (CRR II/CRD V) and resolution 2017/2401. Amendments of (BRRD/SRM). 2019. Regulation (EU) No 575/2013 on prudential requirements for credit [5] European Banking Authority. institutions and investment firms. Final draft RTS on the materiality 2017. threshold for credit obligations . 2016. [6] European Banking Authority. [13] European Parliament and Guidelines on the application of the European Council. Regulation definition of default under Article 178 2017/2402. Laying down a general of Regulation (EU) No 575/2013. framework for securitisation and 2016. creating a specific framework for simple, transparent and standardised [7] European Banking Authority. securitisation, and amending Interactive Single Rulebook: Capital Directives 2009/65/EC, 2009/138/EC Requirements Directive and the and 2011/61/EU and Regulations Capital Requirements Regulation. (EC) No 1060/2009 and (EU) No 2013. 648/2012. 2017. [8] European Banking Authority. Opinion of the European Banking [14] Quagliariello, M. Are stress tests Authority on measures in accordance beauty contests? EBA Staff Paper with Article 458 of Regulation (EU) Series. 2019. No 575/2013. 2018.

www.iasonltd.com 26 Credit Risk

NPL Classiﬁcation A Random Forest Approach Credit Risk

About the Author

Massimiliano Zanoni: Senior Manager He has over 20 years’ experience in the management of complex projects in credit and ﬁnancial risk within leading national and international ﬁnancial services; maturing a combination of quantitative analysis and problem- solving capabilities. Besides, his consulting experience ranging from model development to the implementation of management reporting system; to compliance and IRB validation in credit risk and ALM project, he has worked for Veneto Banca as the Head of Methodology and Reporting - Risk Management Depart- ment. In the same period, he was also the Risk Management delegate in the Risk committees of foreign banks and the Risk Management delegate in the Board of Banca Apulia where he directly oversaw the transition process to the Solvency II framework. In Iason, he has been involved in the development of credit risk solutions and the implementation of a Dynamic Balance Sheet framework.

This article was written in collaboration with Alessandro Palmisano, Milica Antonijevic, Andrea Fenu and Riccardo Redaelli who at the time were working for Iason Consulting.

www.iasonltd.com 28 Argo Magazine

NPL Classiﬁcation A Random Forest Approach aaaa

Milica Antonijevic Andrea Fenu Alessandro Palmisano Riccardo Redaelli Massimiliano Zanoni

rtificial Intelligence has quickly entered in the financial services industry, covering a wide range of applications. This work proposes a structured statistical approach to classify NPL assets according A to their potential recovery level, within an unsecured commercial portfolio. Asset classification is based on the information provided with the NPL portfolios and, possibly, some information gathered during the recovery process. The framework adopted is based on two different components: one targeting the cases that will be recovered and one estimating their recovery level, in particular the work compares a Machine Learning technique known as Random Forest to a better-known Logit approach. The first is introduced with a review of the underlining Decision Tree theory, including its performance metrics, to which an extended Confusion table is added to facilitate the comparison of the event recovery forecasts provided by the different models. The comparison shows that the Random Forest approach is as reliable and performing as the more known Logistic approach, providing a solid overall performance even with a limited set of information. It also successfully tests the ability to compare a portfolio under management to a new one of the same type.

earning is a process meant to identify price estimates and an efficient workout are key patterns and rules within available data, to profitability. L in order to solve unknown problems. It is worth noticing that, though forecasting Traditional modelling performs this task by pre- occurs at single deal level, what really matters suming an underling functional form based on to the investor, is the ability to correctly infer past data to estimate the key parameters for the the amounts which can be recovered at aggre- functions, through a procedure called fitting. gate levels (e.g. at segment level or in different Statistical learning instead is a set of method- geographical areas). ologies where no functional form is assumed and event classification is based directly on the This paper starts by reviewing the theory at characteristics of events; the learning event en- the base of decision trees and its evolution into tails the ability to infer the inputs-outputs map- the Random Forest (RF) approach, showing how ping. The process, to which Machine Learning it can classify inputs into homogeneous pools. belongs, includes two different approaches: su- A brief mention to how the problem of over- pervised and unsupervised learning. fitting is successfully handled, completes the In supervised learning the rule-finding process description of the framework and an overview is based on a set of input-output pairs, while of performance indicators concludes Chapter unsupervised learning classifies events based on “Tree-Based Clustering”. These metrics provide their intrinsic characteristics, i.e. without any the necessary tools to compare models and se- “label” (no output available). lect the most relevant. In the case study considered in this article, Chapter “Dataset Description” is dedicated the training dataset consists of short-term non- to the description of the dataset of NPL expo- performing commercial exposures to be recov- sures, through an overview of its variables and ered by a servicer, which usually purchases such characteristics, focusing both on the static snap- portfolios at a discount and then works them shot and on its payment flow, in order to enquire out to some degree of recovery, hence correct the recovery process.

Issue n. 17 / 2019 29 Credit Risk

FIGURE 3: Tree-Based vs. Linear approach

The modelling approach is presented in To predict outcomes from input variables Chapter “Development Approach”, where the the linear approach relies on a given under- different components are described and com- lying law, potentially “linearized”, while the pared. In particular, the Random Forest ap- tree-based approach relies on the direct charac- proach for the probability of recovery is com- teristics of the training set in order to define the pared to a classical generalized linear regression splitting rules, targeting a given level of homo- model and the recovery rate model is outlined. geneity/granularity. The splitting rules in the Finally, Chapter “Results and Conclusions” output space, obtained by the iterative cluster- presents the results at aggregate and portfolio ing process, are non-linear and non-symmetric. level, showing that the Random Forest frame- As shown in Figure 3, in such a framework work is as reliable as the better-known Logistic the functional approach (left) fails to isolate ho- approach. In addition, a specific test shows that, mogeneous events due to its strong dependence besides evaluating absolute recoveries, the ap- on assumptions regarding the underlying data. proach presented can be used to estimate the On the other hand, the tree-branching approach relative performance of similar portfolios. is able to partially identify existing patterns The work is concluded by suggesting possi- without the need to induce a functional link, ble enrichments of the dataset and evolutions of even when the dynamic relationships between the analysis which would provide useful met- outputs and their determinants extrapolated rics and guidelines to decision making during from the data during the training process are the workout process. blurring. Decision trees used for regression and classification have a number of advantages over the Tree-Based Clustering more classical approaches, as they: • Embrace a learning process and provide a Event classification requires to segment the pre- robust estimate without cross-validation; dictor space into a number of smaller regions • where the output can be considered homoge- Can be represented graphically, are easy neous. to explain and to convert into classifica- The Random Forest is a Machine Learning tions rules; approach, which sets binary rules used to clus- • Mirror human decision-making more ter the event space by recursive branching rep- closely than the regression and classifi- resented as trees. A clustering rule is a set of cation approaches do; splitting points in each dimension of the event • Can easily handle qualitative predictors space, each dividing the set into two separate without the need to create dummy vari- areas on the basis of a well-defined “distance ables; from the mean”. The different sets of rules are all together called the forest. The final model is • Naturally fill missing data through prox- obtained by averaging all the decision trees (rule imity forecasts; sets) composing the forest. A decision trees can • Define automatically homogeneous clus- be applied to both continuous (regression) and ters; categorical (classification) output variables. The main difference between a functional • Easily measure the importance of each approach and a tree-based clustering analysis is variable, and reduce automatically over- illustrated in Figure 3. fitting;

www.iasonltd.com 30 Argo Magazine

FIGURE 4: Splitting process

but, unfortunately, singles trees also have – Randomize the selection of variables drawbacks: at each split (this practice is meant to reduce the dependence of the model • Trees generally do not have the same level upon the most influential variables). of predictive accuracy as some of the clas- • Split the variable space into two by select- sical regression approaches; ing the split-point (determined by the m • They can be quite non-robust, i.e. a previously selected variables) that creates small change in the data can cause a large the ’purest’ sub-nodes. This is the point change in the final estimated tree. The last minimizing the residual sum of squares of issue is overcome by averaging many trees predictions in the specific tree. on the same data set (forest). Formally, for a randomly selected feature j, two regions are defined by the cut-off The last issue is overcome by averaging point c such that: many trees on the same data set (forest). R1(j, c) = (X | Xj < c) (13)

Tree-Splitting Mechanics and R (j, c) = (X | X ≥ c) (14) Prior to proceed with Forest training estimates, 2 j for cross-validation purposes, the dataset is split and find the value of the cut-off such into two separate samples. The larger subset that the following equation is minimized (the training one) is typically used for model (Hastie 2009): development, while the smaller is dedicated to 2 2 testing. The training algorithm can be summa- ∑ (yi − yˆr1) + ∑ (yi − yˆr2) rized with the following steps: i:xieR1(j,c) i:xieR2(j,c) (15) • From the training dataset, draw n boot- • y 14 In each tree obtain a prediction ˆ (terminal strapped subsamples to which the proce- nodes) for each observation by applying dure will associate an equivalent number the clustering rules defined. of decision trees. • Take the arithmetic average of the pre- • Build each tree in the sample by recur- dicted final values, for each observation, sively splitting the set, until the minimum over all the trees defined. size is reached by: • Finally, cross-validate model’s perfor- – Randomly select m variables from all mance on the left-out observations (i.e. the the available p descriptive variables out-of-bag observations, OOB). in the dataset (the procedure is also First of all, this procedure allows to reduce called feature bagging); the bias by growing trees to the maximum 14The bootstrapping is a resampling technique which creates the desired number of samples of the same size and distribution as the original sample, using sampling with replacement.

Issue n. 17 / 2019 31 Credit Risk

depth, i.e. until the minimum node size of 1 the model may end up mapping also the noise is reached. existing in the data, eventually making the fore- The tree-building process selects the best cast less reliable on new datasets, a clear sign of split at each particular step and not the one that overfitting due to correlated variables. will optimize tree in some future step down. If all input variables were considered at each It is important to underline that the extent to node split, the benefits of random variable selec- which the trees grasp patterns in the data is af- tion would get lost, since in every tree the most fected not only by the number of variables tried informative variables would drive the split to at each split, but also by the depth to which trees the same final results. In other words, as for are grown. Therefore, the model is character- multi-linear models, the most relevant variables ized both by the number of variables involved would tend to hide the less important ones in at each branching as well as by the number of the ensemble, as shown in (Hastie 2009) and processed layers (i.e. the depth of the model). (Hastie 2013). Eventually the procedure ends up drawing a The Random Forests approach automatically number of boxes, called terminal nodes or leaves, reduces this risk by randomly selecting, at each identified by clear values of the variable. The split, a subset of explicative variables from the output in each box will be unique, indepen- whole set. The continuous change of variables dently of the number of events included. As at each split limits the chance to systematically mentioned, the final clusters depend on the spe- include correlated variables in the learning pro- cific subset to which the tree-splitting process is cess. The lower the number of variables selected, applied, hence the classification of a given event the lower the correlation risk and the less the maybe different in different trees. model will induce overfitting. What determines the depth of trees is the The natural process of averaging the pre- terminal node size, i.e. the minimum number diction of each single branching in the forest of observations that are contained in the final provided by all trees, reduces the noise picked node. up by each individual tree, hence the sensitivity By selecting many of such subsets and aver- of the model outcome on the initial set selected. aging the results of different tree-splitting pro- Beside the random selection of variables cesses, the model outputs become more stable to reduce over-fitting, tree-simplification pro- and precise, as averaging contributes also to cedures are also used with the aim to optimize the reduction of the model variance given by the model targeting well defined statistical cri- (Hastie 2009): teria or information measures. ρσ2 + (1 − ρ)/nσ2 (16) Pruning is one of the most popular techniques to limit trees complexity and reduce the While the second term of the equation fades overfitting problem. By removing unnecessary away as the number of trees (n) increases, the subtrees from a given level of the tree, pruning first term in the variance equation is minimized locally forget options judged not useful to learn. by feature bagging, which is the random selection These techniques work a-posteriori by removing of variables at each node, reducing the possibil- those subtrees deemed to be irrelevant accord- ity of building highly correlated trees (ρ). ing to a given estimate of error and to grow a The following Chapter will explore in more limited number of branches. details some of the features of the procedure. Data reduction techniques present an alternative for model improvement. These differ from pruning due to their simplification strat- Reducing Overfitting of Tree-Branching egy: while pruning algorithms directly control The quality of a decision tree is usually evalu- tree size, data reduction techniques simplify ated through its complexity and the ability to tree development by pre-processing the dataset generalize rules on different datasets. In the prior to building the decision trees (i.e. reduce attempt to obtain better performance, often cor- the dataset by removing irrelevant features from related variables are added to the tree, introduc- data). This improves the quality of the learning ing over-fitting, a problem that tree-branching process measured by metrics such as reduced- modelling has in common with multi-linear re- error, minimum description length and cost- gressions beside increasing the complexity of complexity pruning. the trees. An effective data reduction technique is Fea- While trying to fit closely a given dataset, ture Selection, which identifies the most rel-

www.iasonltd.com 32 Argo Magazine evant features. This approach is particularly Formally, the Mean of Squared Residuals is suited for problems with few potential out- computed according as follows: comes dependent on a large number of descrip- ∑n (y − yˆ )2 tive variables. MSE = i=1 i i (18) n Performance Indicators where i = 1, ..., n are the OOB observations. Performance measures are key to drive model For instance, if 1,000 trees are grown in one validation and selection and this is no different simulation, each observations of the dataset for the RF approach, for which different metrics would be an out-of-bag observation for 368 trees are used. on average, since the estimating process leaves Usually, for each tree, one third of observa- out an average of 36.8% events randomly se- tions are set aside for validation purposes; such lected to form the test sample. The out-of-bag set is named the “out-of-bag” partition (OBB). MSE represents the prediction error on the test A performance measure frequently adopted set. is the pseudoR2, or Percentage Variance explained In case of a binary target variable, it is pos- (PVE). Calculated using the out-of-bag parti- sible to convert the variable into a continuous tion, it is a measure of the ability to generalize; one by considering the probability of having a intuitively speaking, it shows how close the pre- success. dictions on new observations get to the real A different performance measure is the Brier variance in the output variable, formally: Score (BS), which coincides with the mean- squared error of the predictions for success pseudoR2(PVE) = probability, as a transformation into the contin- ∑n (y −y)2 ∑n (y −yˆ )2 uous domain of Boolean target variable. Since i=1 i − i=1 i i MSE (17) n n its value is in the range (0, 1), it can be used as n 2 = 1 − ∑i=1(yi−y) Var(y) n an absolute measure of model performance. Formally, it is given by the following for- where yî and y are the average of real and es- mula: timated outputs respectively and MSE is the sum of squared residuals divided by the sample ∑n (qˆ − y )2 BS = i=1 i i (19) size and Var(y) is the variance of the response n variable. where yi is the observed outcome (Boolean pos- A small pseudoR2 value generally indicates itive/negative encoded with 1 and 0) while qî is that the model is sub-optimal, while high val- the predicted probability of success (a number ues, especially if higher than 0.6, confirms that between 0 and 1). the model represents a good fit of the given A more complete performance measure for dataset. binary variables is the ROC curve (Receiver Op- Clearly pseudoR2 can take also negative val- erating Characteristic)15, a popular measure that ues in case that the ratio between MSE and provides information on the correct classifica- Var(y) is bigger than 1 or, in other words, if tion of positive and negative events based on the variance of predictions is greater than the the two following dimensions: variance of the real values. The perfect model (i.e. the model which • The Sensitivity (on the vertical axis) is the gives predictions equal to the observed values) True Positive Rate (TPR), i.e. the fraction 2 would be associated to pseudoR equal to 1, of success cases (e.g. recovery) correctly which can be used as the theoretical limit for identified; models’ performance. Another common performance measure for • The Specificity (on the horizontal axis), RF models is the Mean of Squared Residuals, cal- is the True Negative Rate (TNR) and pro- culated as the mean squared error over all OOB vides the fraction of observations correctly observations. tagged as failure (e.g. default, failed recov- This measure therefore provides a proxy of ery). Hence the quantity 1−Specificity is the error rate of the model on new data, it is the fraction of negative events incorrectly in the range (0, ∞) (the smaller the better for classified as positive (i.e. the False Positive performance). Rate, or FPR).

15The name originates from Communication theory where it was initially used.

Issue n. 17 / 2019 33 Credit Risk

TABLE 5: Confusion Table

TABLE 6: Confusion Table with totals

In general, the recognition of events in produces only 2.2%, offering a rough measure model application requires the definition of a of the model’s ability. In addition, the Extended cut-off value that separates positive from nega- table includes two more measures: tive cases. Referring to 1−Specificity, instead of Specificity alone, the ROC curve is just the plot • The fraction of correctly forecasted success of TPR against FPR for different cut-off values. and fail cases (OK and OK on sample) and Following Table 5 represented below, the their complementary mistake ratio (KO); two measures are defined as: • The fraction of correctly forecasted fail and success events in the real samples. A Specificity = : P(Q = 0 | Y = 0) (20) A + B Table 7 can be read as follows: the model identifies correctly 20,523 out of the 21,047 negative events (Not recovered), i.e. 97.8% of the portfolio, D but the real number of cases with no recovery Sensitivity = : P(Q = 1 | Y = 1) (21) C + D (real negatives) is only 20,877, hence the hit ratio of the model in the Not recovered subset is actu- where Q is the forecasted outcome, while Y ally 98.3%. This dynamic reverses for positive is the observed one. The two measures distin- cases, where the model hit ratio of Recoveries is guish between different types of mistakes. only 19% with reference to real recovery events. The perfect model, i.e. the one forecasting The overall hit ratio of negative and posi- all outcomes correctly, is obtained when both B tive cases is nearly 96%, registered in All-True, a and C are equal to 0, i.e. when Sensitivity = 1 measure of overall performance that, however, and Specificity = 1. does not distinguish mistakes of different types. In the present work, positive cases consist This distinction, however, may be vital in cases of relevant recoveries and once the table above where the cost of different mistakes is not ho- is filled with the real and predicted values, the mogeneous (as for credit granting for instance, relative Confusion table is obtained (see Table where landing money to the wrong customer 6), where the totals are added for completeness. may lead to the loss of large amounts of capi- The example shows the number of success tal).16 and fail cases during model application on the An additional way to measure the overall whole population, but absolute figures may not performance of such a model at sight is repre- be the best way to compare two different models sented by the Area under the ROC Curve (AUC). or cut-offs. Hence, in this work, the following The bigger the AUC, the closer the curve top extended Confusion Table will be used (see Ta- will be to the top-left point of the chart, repre- ble 7) where, beside absolute values of success senting the perfect model. Values around 0.8 and fail (real and fitted), their portfolio share is are usually considered very good. An example reported. of ROC Curve is shown in Figure 5. In the example reported, while the portfolio A key feature of the graphical representation includes 3% of recovery events, the forecasted of ROC is that it can plot several models in the 16Similarly, in the recovery process, reducing the effort on a false negative (i.e. a dossier erroneously tagged as non recoverable) will reduce recovery flows, while indulging on a false positive (i.e. a dossier erroneously tagged as recoverable) may increase operation cost only marginally.

www.iasonltd.com 34 Argo Magazine

TABLE 7: Extended Confusion Table - Example same chart, making it easier to compare their it to that of other variables. performances. Another important feature of the Random An alternative to Mean decrease in Accuracy is Forest modelling framework is the possibility to Mean decrease in node impurity, proposed in measure the contribution of each single variable. (Breiman 2001). This evaluates the importance This is obtained through two specific measures: of each variable by comparing the “purity of Mean decrease in Accuracy and Mean decrease in the node split” in the alternative of including or node impurity. excluding the given variable from the set used The Mean decrease in Accuracy evaluates the for splitting. importance of each variable by analysing the The impurity is defined as a Residual Sum of change in the prediction error due to the change Squares (RSS): in values of a given variable. The test is per- 1 Imp(X ) = p(t) ∆i(S , t) formed as follows: m n ∑ ∑ t T teT:V(St)=Xm 1. The estimated model is used on test ob- (22) servations and the MSE is calculated on where: these predictions; • ∆i(St, t) is the impurity decrease (mea- 2. The value of one variable at a time is ran- sured as a decrease in the RSS) due to domly modified, then the model is ap- the new split based on the given variable plied again (ceteris paribus) obtaining new Xm; predictions and recording a new MSE for • ( ) each variable in the model; The weight p t is the portion of trees n reaching the node t over all trees in the for- 3. The difference between the two MSEs est. Hence, it is the probability of reaching (MSE of the original data and the MSE of the node t. the case with value changed) is recorded and averaged over all trees and for every Intuitively, by excluding the most relevant vari- variable; able there will be a higher increase in RSS, which gives us a very good approximation of 4. Finally, the distribution of MSE differ- the variable’s importance for the model accu- ences is normalized in order to compare racy.

Issue n. 17 / 2019 35 Credit Risk

FIGURE 5: ROC curve

Dataset Description In details, the retail portfolio to be analysed con- tains 22,290 observations, 42 of which must be The dataset analysed in the exercise includes a dropped since the fiscal code is missing and key list of non-performing retail exposures related information cannot be assigned, even if the re- to unpaid utility bills with few information covery rate for this subset is significantly higher about the customer (listed in the Table below) than the portfolio average. No further explana- and limited information about the exposure. In- tion is provided for this pattern of the excluded formation is provided both as snapshot at a observations. given date with geographic and customer details associated (static info dataset), and as a flow Data Quality and Filters of payments at each date (dynamic info or cash The distribution of counterparties in the dataset flow dataset). is represented by debtor age in Figure 8. This value has been capped at 95, because counter- Static Info Dataset parties with age over 95 years were considered Table 8 shows a summary of the information as an anomaly. Also records with age up to 118 provided in the first dataset. years are listed, hence the relative records were The payment flow dataset presents several dropped from the development dataset for data data quality issues; for instance, the total re- quality reasons. covery amount should be equal to the sum of As known, recoveries from non-performing flows for a given dossier, but often this is not the loans are distributed mostly in bimodal mode, case. For analysis purposes, when the two are i.e. most dossier in the portfolio present either different, the sum of flows is taken instead of no recovery or a high recovery rate, leaving only the recovery amount provided in the portfolio a fraction in the middle of the recovery scale, as snapshot. summarized in Figure 7.

www.iasonltd.com 36 Argo Magazine

FIGURE 6: Counterparty distribution by debtor age

TABLE 8: Observations by Recovery level

Dossiers with a very small recovery level, covery process. Intuitively, this is linked to the both in relative (recovery rate) and in absolute duration of the process and, as shown in the terms (e), might introduce noise in the esti- following histogram, it affects the average re- mation process, hence a filter was outlined to covery rate too, at least up to a certain number exclude them from the development dataset. of contacts. On average, dossiers with 140-170 This filter is set to consider as relevant only contacts present a 50-55% recovery level, but those dossiers for which the recovery rate is increasing the number of contacts does not im- above 10% of total exposure and, in any case, prove the rate of recovery, which instead drops the amount recovered exceeds the threshold of significantly to around 30-35% on average, as 50 e, which finally excludes 204 dossier, leav- shown in Figure 9. ing 647 observations in the recovery list out of 21,524. A possible explanation for this behaviour is Most excluded dossiers (over 65%) are between that a significant effort is made on cases yield- 750 and 2,500 ein size, but show a similar recov- ing a poor and slower recovery, probably in the ery level around 3.5%, slightly above the overall hope of reaching a target collection level. In portfolio average of 2%, confirming their border- principle, this represents an area of improve- line status of “irrelevant recovery”, at least for ments that might take advantage of statistically modelling purposes. A graphical description of set benchmarks. these dossiers is given in Figure 8. A more detailed evaluation of these cases is pro- The variable Contacts represents the number vided in the next Section, based on payment of interactions with the counterparty in the re- flow analysis.

Issue n. 17 / 2019 37 Credit Risk

FIGURE 7: Number of observations by Recovery level

FIGURE 8: Analysis of excluded dossiers

Cash Flow Dataset at each duration, showing that a large majority (over 70%) is completed within 6 months, As mentioned earlier, besides the dataset with though with less success in terms of recovery portfolio and customer information, the “snap- rate. shot”, a second dataset with information on recovery flows, was provided. Each payment It is clear that the dynamic of the recov- is marked with the payment date, allowing ery process offers important feedbacks on the to build the precise recovery history for each effectiveness of the recovery process, consid- dossier. However, this dataset also presents a ered that there is a large number of dossiers number of quality issues: with very short collection periods. In fact, over • For a number of exposure, the sum of all 450 dossiers end their collection process within flows in this file did not match the total re- the first month and the recovery rate is loosely covered of the first dataset (in these cases, linked to the total duration of the collection pro- the total of the flow data set was consid- cess, as well as to the number of contacts, as ered instead); mentioned in the previous paragraphs. • Often the list of dates relative to the different flows does not present a regular Recovery Process collection process with an even distribution of payments, but a number of equal In any dossier, the cumulated recovery, due to payments executed on the same date; successive cash flows, increases with time up to the final recovery rate provided in the snapshot • In few cases, beside the list of equal pay- dataset. ments totalling the exposure to recover, This dynamic may provide important informa- also a bulk payment appears at the end of tion to collectors. In fact, different dossiers ex- the period, equal or close to the sum of the hibit different collection trends, many ending other small payments. In some cases the the collection process in one single flow/month, bulk payment was excluded, as it was con- as shown in Figure 10, while others continue to sidered a double entry. 21 cases where a collect for dozens of periods, reaching on aver- clear fix was not available were completely age a higher recovery rate. excluded from the dataset. Along the whole recovery process, at any given The histogram in Figure 10 counts the number period T, the Average Cumulated Recovery Rate of dossiers completing the collection process (ACRR) is given by the sum of payments, di-

www.iasonltd.com 38 Argo Magazine

FIGURE 9: Average Recovery Rate by contacts vided by the total exposure of dossiers, which following the previous notation). The separa- collects any amount until T, including those tion of the respective lines represents a measure that completed the collection process at an early of the non-homogeneity of recovery dynamics stage: of different dossiers. Figure 12 shows the first and last decile (10%) ∑t,T R ft for the recovery processes in the D subset, rep- ACRRT = (23) 3 ∑T ET resented by the red curve in Figure 11. Clearly, Given the different recovery dynamics, the the top decile is characterized by a quicker pro- definition of the average recovery curve should cess and a much higher recovery rate, while the be based on processes with homogeneous dura- last one reaches only a marginal recovery rate. tions. In the following graphs, this is obtained This analysis supports the investigation of by restricting the recovery process to dossiers the behaviour presented by Figure 9, i.e. the with a recovery duration above: decreasing effectiveness of clients’ contacts. It is useful to group dossiers with a longer re- ∑ > R ft,K t,T,K D (24) covery process in two different sets: ∑T>D ET,K • Effective processes, made of 65 dossiers (to- with K selecting those dossiers with a collection talling an exposure of 101,540 e) which process lasting or exceeding duration D. underwent a number of contacts between Hence, with D = 1 (named also D ), the 1 86 and 172; set includes the whole portfolio (marked with a blue line in Figure 11); considering all the • Slow processes, composed of 22 dossiers quickly recovered dossiers, the curve is steeper (totalling an exposure of 39,206 e) which in the initial part. Clearly, the large number underwent a number of contacts above of quickly resolving dossiers creates a quick re- 172; covery boost which however reaches a lower recovery level with respect to a subset including and to compare their recovery dynamic (see dossier with a slower recovery process lasting Figure 13), as well as their cumulated recovery at least 3 months (market with a red line). This level through time. It is clear that not only the also reflects the evidence of Figure 11, where effective processes subset reaches a higher recov- longer recovery process yields higher recovery ery rate, but it also reaches it quicker, collecting rates. Of course this effect would be more a bigger portion than the other subset in nearly marked excluding dossiers with a higher col- every period. lection period (D > 3). In Figure 13, the horizontal axis represents Of course, each curve is an average of many the number of months in the recovery process, dossiers with different recovery dynamics and while the vertical one represents both the por- durations. A possible way to show this variance tion recovered in that particular month (left is to plot the first and the last 10% of dossiers in scale) and the cumulated total recovery on the the recovery process within a given subset (Dx, subset (right scale).

Issue n. 17 / 2019 39 Credit Risk

FIGURE 10: Number of dossiers and average Recovery Rate by process duration

FIGURE 11: Average cumulated Recovery Rate dynamic for different portions of portfolio

Development Approach approach similar to LGD modelling is useful, in order to overcome the problem of extreme dilution of recovery parameters by non recovery This Chapter illustrates the methodology events. adopted to estimate and test the two different To this extent the estimation process is di- models used to forecast recoveries in a portfolio vided into two parts: the first one meant to of NPL exposures: a model dedicated to fore- estimate the Probability of Recovery (PR), i.e. the cast the probability of recovery of the single probability that a “relevant” recovery can occur dossier and one to estimate the recovery rate in for a given dossier; while the second model is case of a recovery. This model is estimated only targeted to estimate the Recovery Rate of those on cases where an effective recovery is obtained. specific exposures where a recovery effectively The model identifying a recovery event is occurs (Recovery Given Recovery, or RGR). Hence estimated with two different approaches: first the amount of effective recovery will be quan- through a logistic/generalized linear approach, tified by this second model, in case a recovery then using Random Forests. This is meant to occurs. provide a classic estimation method in addition The two models are estimated and tested to the Machine Learning approach in order to separately on the same variable set. They will offer a known benchmark to compare the result thus be statistically independent of each other, of the new tteblack box approach. but dependent on the availability and quality of Depending on the asset type, the number of the same data and potentially on overlapping NPL with zero or tteirrelevant recovery can be explanatory variables. very high (as in the present dataset). Hence, an

www.iasonltd.com 40 Argo Magazine

FIGURE 12: Average cumulated Recovery Rate dynamic for processes over 2 months

FIGURE 13: Comparison of dossier Recovery

In order to provide a frame for comparison, above the cut-off C that identifies a probable beside the Random Forest approach, the two recovery: models are estimated using also a generalized ( linear approach, i.e. a logit approach for the 1 if RPi ≥ C Ii = (26) probability of recovery and a generalized multi- 0 if RPi < C linear one to model the recovery rate. The “generalized” feature here refers to the fact that This procedure represents a non-linear trans- both quantitative (continuous) and qualitative formation of the probability of recovery as- variables are considered. signed by the model (logistic or RF one), which otherwise, on average, determines the number of recovered dossiers N (i.e. it is calibrated to The Total Recovery (TR) is the sum of the rec the recovered sample): amounts recovered from each exposure with a relevant recovery: ∼ RP · N = Nrec (27) where N is the number of dossiers in the port- N N folio. TR = ∑ Ri · Ii = ∑ RGRi · Ei · Ii (25) i=0 i=0 Depending on the definition of the threshold, the number of effective recoveries may higher where Ri is the recovered amount, Ei is the total or lower that average or, specifically, in line with exposure, RGRi is the recovery rate considering the real number: only positive events and Ii is a flag identifying the cases with a relevant recovery. The flag is N RP · N =∼ I (28) assigned through the probability of recovery ∑ i i=0 model and a threshold defined on the given distribution, i.e. when the recovery probability is For the same reason, the average recovery

Issue n. 17 / 2019 41 Credit Risk

amount, in general, is different from the esti- elled in both the multi-linear and RF approach mated recovery: to provide a comparison frame with which most readers may be conﬁdent. N N ∑ RGRi · Ei · RPi 6= ∑ RGRi · Ei · Ii (29) i=0 i=0 Threshold Deﬁnition

An additional benefit of the independence Two different thresholds are considered during of the two models is that, even when the accu- the analysis and modelling of the dataset: racy at deal level is inadequate, at portfolio or • main aggregated level, the accuracy of recovery The relevance threshold, meant to identify estimates may be much better, confirming that which dossiers in the dataset should be the approach is quite robust. considered as recovered, i.e. those for Each of the two models estimated in a given which the recovered amount is sufficient framework will be characterized by specific per- to be considered a successful case; formance levels, measured by a different set of • A cut-off in the estimated probability of re- indicators. An independent error is estimated covery, assigned by the first model and into the final recovery estimates. Since the two used to decide which of the dossier is to models are statistically independent, their er- be counted as recovered in the application rors may partially cancel out at an aggregated phase. level, thus potentially providing an estimate of effective recovery more accurate than the per- The relevance threshold is introduced to ex- formance of each model would assure. clude from the development dataset minor re- The performance measure for model esti- covery events, which would otherwise be con- mated with the linear approach will be a classic sidered as successful events and might intro- R2 correlation coefficient, in the adjusted ver- duce further noise both in the rate of recovery sion for the multi-linear RGR model. A ROC parameter, representing a marginal component curve will also be used to visually show the of the portfolio, and in the probability of re- quality of the PR model in both approaches. covery, where they would be counted as non The performance of the models estimated with recovered. the Random Forest approach will be evaluated This threshold is defined on a judgemental in terms of Percentage of Variance Explained (PVE) basis, after analysing the characteristics of the and Increased Node Purity. recovered dossiers (evidence is given in Chap- ter “Dataset Description”). It is based both on Probability of Recovery Model a relative recovery level (the recovered amount should be at least 10% of exposure) and an ab- The probability of recovery is the value assigned solute recovered amount (it should not be lower by the specific model to each exposure repre- than 50 e); below these two levels the dossier is senting the probability that it may yield a re- not considered a recovery event. covery above the defined threshold. From that, An estimated case of recovery occurs when it is possible to determine if the dossier is ac- the probability of recovery assigned by the rela- tually recovered. In practice, this is the value tive model exceeds the given cut-off, hence the triggering a recovery event. recovery event depends specifically on the cut- To provide a term of comparison for the off value defined. Increasing the cut-off will probability of default, the model estimated with increase the count of recovery events marked as Random Forest will be compared to a different non recoveries. model estimated with a logistic approach. This, The cut-off in the estimated probability of re- being a well-known and consolidated method- covery can be defined in different ways after a ology, provides a solid benchmark. model is estimated, in order to determine which On the other hand, the probability of recov- dossier should be considered a successful recovery estimated in the Random Forest approach ery. can also be transformed in a list of recovery events, through a predefined threshold. Hence, A) A possible value is obtained by dividing that output can be compared directly to the lin- the dataset in real recovery and non recov- ear approach through the confusion matrix and ery events and then plot each one along the correlation coefficient. the probability of recovery assigned by the Also the RGR model component will be mod- model. The two histograms will distribute

www.iasonltd.com 42 Argo Magazine

FIGURE 14: Quantile distribution of Recoveries (red) and Non Recoveries (pink)

TABLE 9: Predicted Recoveries versus Cut-Off Probability

along the horizontal axis designing two dis- previously, but even this small difference tinct distributions: non recoveries on the left identifies 137 more recovery events. and recoveries to the right. The cut-off can It is interesting to notice that the first two be taken as the approximate value at the in- methods identify the same cut-off, though in tersection point of these two distributions. the first case is based on a graphical-manual Figure 14 represents quite well the polariza- estimate. tion of non-recoveries occurring in the NPL Obviously, estimating more recovery events portfolio in analysis: nearly all are included to get closer to the real number does not in the first bar to the left. Unfortunately, re- imply that the model is actually identifying covery events are evenly spread along the them correctly, hence the performance of the probability parameter, apart from a bunch of model as a whole may even decrease. dossiers with a high probability of recovery. This makes the distribution of recoveries not C) A totally analytic approach is to define the well defined and the cut-off can be taken cut-off that maximizes the overall number after the non recovery bar, i.e. at 0.1. of success cases - both negative and positive - by finding the value that: B) Reducing the cut-off identifies more recovery events; an alternative and more quanti- • Minimizes the difference TNR – TPR tative method to define the cut-off is to set considering positive values only; it in a way that the number of estimated re- • Maximizes the sum between TNR and coveries equals that of real recoveries, or at TPR (which is commonly known as least as close as possible. The Table 9 shows Youden’s index). the number of predicted recoveries as a function of the defined cut-off in a portfolio with In the following paragraph, the three meth- 626 real recoveries. This approach suggests ods will be compared by evaluating model per- a cut-off of 0.09, very close to the 0.1 se- formance through a common tool, the Extended lected with the visual approach presented Confusion table.

Issue n. 17 / 2019 43 Credit Risk

TABLE 10: Logistic model for the Probability of Default

Linear Approach B) The Confusion Table 12 based on a cut-off probability equal or above 10% defined by This paragraph presents the best logistic model method B produces a portfolio mix very attained on available information and compares similar to the real one (3.1% recovery vs the output of different cut-off levels to real re- 2.9% real) but with a slightly lower per- coveries. formance on recovery dossiers since most The final dataset for model estimate is a port- of the increase in recovery signal, with re- folio including only dossiers with a fiscal code spect to the previous cut-off, is actually and counterparties with age up to 95 years but wrong. excluding, as explained previously, recoveries below 50 eor with recovery rate below 10%. C) The Confusion Table 12 based on a cut-off An R generalized logistic function is used to probability of 3% defined by method C, model recovery events. The best result includes choosing the value that grants the same the variables listed in Table 10, showing their performance of both positive and negative statistical relevance. forecasts on the real sample (OK on sample). This method identifies a much larger In Figure 15, the ROC Curve associated to number of recovery events (67.3% vs 22% this model visually shows how much the model of the previous methods) but at the ex- improves the casual guess (represented by the penses of a huge increase in the recovery straight line) and, at the same time, how far signals (7150, or 33,7% of portfolio), most it still is from the ideal model, represented by of which wrong (94.1%). upper right hand side angle. Though the ROC does not depend on a spe- At the base of this approach, granting the cific cut-off, the effective performance of this same performance for positive and negative model, in terms of correctly identified recovery events, is not only the assumption that the same events, depends on the cut-off chosen, whose cost is associated to both types of mistakes – effects can be analyzed in details through the Ex- not plausible - but also that the two event types tended Confusion matrices shown below. Each cover, more-or-less, the same portion of the pop- criterion listed in the previous paragraph is ulation, which is clearly not the case. based on a different cut-off value, defined by The first two methods identify very close the three methods described. cut-off values and though the first presents a higher all-true performance rate (95.8% vs of A) For completion the higher cut-off is also 95.3%), the second selects a number of recovery analyzed besides that identified by the cases closer to the real one (655 instead of 512 vs numerical-graphical approach.The Confu- 626 real), thus forecasting a portfolio mix more sion Table 11 based on a Cut-off probabil- similar to the effective mix (3.1% instead of 2.4% ity equal or above 11% defined by method vs 2.9% real). A which offers a slightly better perfor- For this reason the second option will be mance on recovered dossiers and overall used to compare the RF to the Logistic approach (OK on sample). in terms of overall effective recovery.

www.iasonltd.com 44 Argo Magazine

FIGURE 15: ROC curve - Logistic approach

TABLE 11: Extended Confusion Table - Cut-off selected with method A

TABLE 12: Extended Confusion Table - Cut-off selected with method B

Issue n. 17 / 2019 45 Credit Risk

TABLE 13: Extended Confusion Table - Cut-off selected with method C

TABLE 14: Extended Confusion Table - Random Forest approach

Random Forest Approach model and final classification thus, as a consequence, the numbers in the Confusion table may After the logistic approach, the same model was be slightly different. estimated through a Random Forest approach As introduced in the initial Section, a Ran- on available information. As for the logistic dom Forest model is quite complex to repre- model, the output is a continuous variable cali- sent visually, consisting in the average of many brated to a probability of recovery, which is then trees, each with several split values for each turned into a recovery event through the cho- explanatory variable. Contrarily to the logistic sen cut-off. The estimated model, with a 10% approach, the variable selection is not neces- cut-off, yields the Extended Confusion shown sary for the Random Forest, where all variables in Table 14. are usually included - and ranked - and the With a percentage of predicted recoveries model is evaluated through its standard perfor- equal to 3.5%, the RF model slightly overesti- mance measures reviewed at the end of Chapter mates positive events with respect to the real set, “Dataset Description”. where the same percentage is 2.9%. On the other For the RF model, two of the most popular hand, the logistic approach underestimates the performance measures are summarized in Ta- recoveries (2.4%). By directly comparing the ble 15: the Increase in node purity and the Mean Confusion table of the logistic model presented increase in Accuracy17. in the previous paragraph, relative to the 10% They identify the number of contacts, the cut-off, the RF model identifies more “True pos- dossier exposure and the duration of the dossier itive” (170 cases against 125), in particular those treatment as most relevant variables. in sample, producing a much better OK on sam- On the other hand, performance and robust- ple (27.2% versus 20%), but this occurs at the ness depend on parameters which need to be expense of producing many more “False posi- properly tuned during the estimation process to tive” (564 against 387). get the best model, specifically: It is worth noticing that, differently from the deterministic framework of the logistic ap- 1. The number of trees to be trained (ntree = proach, the Random Forest approach includes 500); a non-deterministic component that introduces minor changes at each new run, hence each 2. The minimum size of terminal nodes in estimate process will yield a slightly different each tree, (nodesize = 50); 17The Mean increase in Accuracy evaluates the importance of each variable by analysing the change in the prediction error due to the change in values of a given variable. As the name suggests, it is strictly related with the Mean decrease measure described in Chapter “Tree-Based Clustering”. The same holds for Increase in node purity.

www.iasonltd.com 46 Argo Magazine

3. The number of variables to be selected eters, but this work focuses more on compar- every time a new tree is trained (mtry = ing Machine Learning to the classic Logistic 2). approach and will not inquire further about the effects of these changes. The choices in the estimated model are ex- The Logistic and the RF recovery probability pressed in brackets above. model can be directly compared through their The literature provides examples of the ef- relative ROC curves, both plotted in Figure 16, fects on the characteristics and the performance which conﬁrms a slightly better performance of of RF models due to changes in these param- the Random Forest (blue line).

TABLE 15: Performance metrics - RF-PR model (cut-off set at 10%)

FIGURE 16: ROC curve - Comparison between RF and Logistic approach

TABLE 16: Extended Confusion Table - Random Forest as a classiﬁer

Issue n. 17 / 2019 47 Credit Risk

Random Forest as Direct Classifier Recovery Rate Model This paragraph introduces the model used to The RF model examined in the previous estimate the recovery rate for those dossiers for paragraph was estimated in “regression mode” which a recovery occurs. It is developed with in order to be able to compare it directly with a standard multilinear approach on the basis the Logistic approach on a similar framework. of information available on relevant recovery However this is not the only mode in which the dossiers. For this model, no Random Forest RF approach can be used: as an alternative, the alternative will be presented as the target vari- RF can be used directly as a classifier on the able does not represent a classification but a basis of the prior distribution. In this case, there continuous value. is no need to define a cut-off for the classifica- The Recovery Rate (RR), which is the target tion of recovery events, since they are directly variable to be estimated, is defined, using the determined by the branching algorithm for the previous notation, in the same way at granular RF procedure. and aggregate level as:

The outcome of the RF estimate in classifica- Recovered Amount RR = (30) tion mode is summarized in Table 16. Exposure Notice that this “classification mode” pro- The geographic variables (related to Macro duces the best results, by far, in terms of correct Region) are maintained in the model even with forecast of recoveries within the specific subset low relevance, as they characterize the final level (67.5% versus the usual 22-24%). However, this of aggregation, as shown in in Table 17. is due to an extremely careful classification of re- The Adjusted R2, equal to 10.4%18, of the covery events, as only 126 of them are identified. selected model confirms that the overall per- This generates a very different portfolio from formance is quite low. However, this does not the real one (0.6% of fitted recoveries against prevent a good forecast of overall recoveries at 2.9% of real ones). aggregate level (i.e. Macro Region one), as it will be shown later. A possible explanation for this Given the small number of recoveries iden- outcome is that the model squeezes recovery tified, this performance is not maintained once forecast around the real average, as shown in the population is considered instead of the refer- Table 18. ence dataset: in this case the hit rate of recover- Dossier size and exposure are consistent ies drops to 13.6%, while that of non recoveries throughout all geographic areas, underlining reaches the top score of 99.8% of negative events that there is no concentration in size in any spe- identified over the whole population. cific area. This means that, when the RF is used as The overall forecast for the Recovery Rate an autonomous classifier, it is very prudent in (RGR) is quite good (62.34% vs 62.6%), but also the identification of recoveries, providing very at geographic area level, with the largest gap good performances, but, in general, it depends occurring for Centre-North estimates where the on the original distribution of positive events. model underestimates recovery rate by 2.3% on average.

TABLE 17: Recovery Rate - Linear model

18On the other hand, the model is consistent at a more technical level: the F-statistic yields a value equal to 13.12, which is bigger than the critical f-value based on 6 and 619 degrees of freedom, which is equal to 1.774. This conﬁrms that the model is statistically signiﬁcant.

www.iasonltd.com 48 Argo Magazine

TABLE 18: Comparison of Recovery Rate outcome

Results and Conclusions This means that the aggregated exposure related to recovered dossiers cannot be calculated The total recovery forecast is the result of simply as the product between exposures and the application of two independent models to predicted recovery probabilities: in fact, no ex- each dossier. Recovery estimates at dossier posure weighting is involved in the recovery level are averaged over the aggregating di- probability, requiring to sum up the specific mension (Macro Region) and then compared to exposures of real or fitted recovered dossiers the amount effectively recovered on the same (non-linear component). On the other hand, the perimeter, in order to measure the final perfor- recovery rate is the exposure-weighted sum of mance of the framework. Before analysing the recovery amounts: output data, it is useful to review the definitions ∑ f it Ri used in the tables and diagrams reported. RGR[ = (35) The Expected Recovery (ER) at granular level ∑ f it Ei is a non-linear function of the estimated prob- Hence, the aggregated recovery amount is ability of recovery multiplied by the exposure given as the product of the average recovery of the given dossier and the estimated recovery rate by the total exposure of estimated recov- rate. ered dossiers: The sum of the expected recoveries within ! ! a given perimeter (Macro Region) yields the to- N Rec = RGR[ I · E = RGR[ E tal recovery at aggregated level to be compared f it ∑ i i ∑ i i=1 f it with the amount effectively recovered within the (36) same perimeter and not from the same specific The results in terms of recovery parameters dossiers. are summarized in Table 19, which shows that N the recovery rate is fitted on average rather ERc = ∑ Ei · Ii · RGR[ i (31) well, while the number of identified recover- i=1 ies is less precise, particularly in some Macro where: Regions: while the logistic approach is more pru- ( dent (i.e. it underestimates), the Random Forest 1 if RPi ≥ C Ii = (32) model overestimates the real rate of recovered 0 if RP < C i dossiers. This quantity is different from the recovery When using the cut-off optimized for the estimated from dossiers effectively recovered: probability of recovery defined by the Logit approach (PR >= 10%) clearly the Random Forest M approach (RF*) overestimates the number of ER = ∑ Ej · RGR[ j (33) j=1 dossier effectively recovered (3.5% vs 2.9% real), since its output is calibrated with a different where the sum is over the M real recovered method. However, if a specific cut-off is defined dossiers. for the RF default probability, based on the same It is worth mentioning that the probability principle which detects a number of recoveries of recovery in a given perimeter is the ratio of close to the real one, the resulting Probability dossiers recovered, real or estimated: of Recovery at macro-region level (RF optimal) gets very close to the real one and presents a 1 N more regular behaviour compared to that of the RP = ∑ Ii (34) N i=1 logistic approach.

Issue n. 17 / 2019 49 Credit Risk

TABLE 19: Fitted vs. Real Recovery parameters summary

TABLE 20: Fitted vs. Real Recovery parameters

In terms of amounts recovered, the model Conclusions output is summarized in the Table 20, comparing the different components. Note that the This work has analyzed the collection profiles expression ’exp · RealPR’ next to real exposure in a portfolio of retail unsecured bills managed of effectively recovered (Real), represents an av- by a servicer, with the purpose of modelling the erage based on the real probability of recovery recovery capabilities on the bases of geographic defined at macro-region level. The two amounts and personal characteristics of the debtors, as differ due to the non-linear nature of the frame- well as those of the assets (very few information work adopted, i.e. the real recovery is the sum on both were available). of amounts effectively recovery while the other The analysis intends to compare two differ- is an average over the entire cluster exposure. ent approaches: an innovative Machine Learn- ing technique known as Random Forest to the Firstly, the Table 20 shows that the majority more classical logistic approach and introduces, of recovery occurs in the North (38,2%) though besides other metrics, extended Confusion ta- the majority of the exposure is concentrated in bles used to compare the different outputs. the Centre South (41%). The work shows that it is possible to develop The exposure of recovered dossiers in any a structured quantitative framework with the model is the sum over dossier recovered accord- aim to classify NPLs into recovery/fail opportu- ing to the specific model and cut-off. If the RF nities, even with limited information. Generally, approach is considered with a cut-off optimized the adoption of statistical analysis promotes the for the Logit model (Exp-RF) the forecasted ex- improvement of data quality and identification posure of recovered dossiers estimated with the of more data sources. RF is less precise than that estimated with the The result is obtained in a framework that e Logit model (922,8 vs 831,7 th ). separates the recognition of recovery events On the other hand, when a cut-off is specif- from the estimate of the recovery amount. This ically defined for the RF (RF Optimal) the es- separation avoids that estimated recovery rates timated exposure gets much closer to the real get too diluted due to the high number of no- one (697,3 th e) with respect to the Logit (748.8 recovery cases. vs 831.7 th e). The two components are estimated sepa- Considering the amounts effectively recov- rately through two distinct models: one identi- ered instead of the dossier exposure, the RF ap- fying the limited set of recovered dossier and proach performs slightly better than the Logit the other targeting the recovery rate in those with both the non-optimal and the optimized dossiers. Both models are estimated on the cut-offs. same variable set. At the same time, the work introduces a machine learning approach called Random For-

www.iasonltd.com 50 Argo Magazine est, yielding a more robust models in an area given stages could be collected to push or where very few assumptions can be made, the stop the collection process, depending on the classification of recovered dossier. Here the RF marginal value of the expected value of that approach was compared to a standard Logistic action. approach to provide results in terms of a known For instance, in a given dossier, 30% of the methodology. exposure was collected in times longer and with As a classification tool, the RF has proved more process effort (i.e. Contacts) than the stan- to be at least as reliable and performing as the dard curve recommends. On the basis of col- Logit approach, but with no need to define spe- lection times and recovery costs of successive cific conditions on explicative variables. In some phases, it may be calculated that the present cases, the RF approach can be used directly as value of expected recovery will not balance, so a classifier to identify recovery cases, without that the process for that dossier will be inter- the need to define any cut-off. Detailed analysis rupted or will follow a cheaper collection chan- shows that the approach presented is suitable to nel. estimate the portfolio value even with limited After an appropriate calibration phase, the information, both in absolute and relative terms, decision making process can be divided into i.e. when changes in value of a new portfolio several streams following specific rules and are estimated relatively to a known portfolio in- collection channels and, in case the number of stead of the overall value. The comparison test dossier justifies it, the service firm can decide was built by selecting the dossiers available into to design a specific collection process to handle two portfolios equal in size and exposure, but otherwise stopped dossiers. with one presenting 25% more real recoveries. The model estimated on the known portfolio was able to recognize the bigger recovery level of the new portfolio, showing that the framework is suitable to be used in relative terms besides absolute terms. Although, in statistical terms, the performance of the model is generally low, mainly due to the very limited information available, at aggregate level the results are quite impressive, even in absolute terms.

Further Developments In order to provide useful metrics and guidelines to support decision making, further possible enhancements can be proposed both on the dataset and the analysis framework. In particular, variables connected to the recovery process could be collected in order to allow more detailed analysis at different stages of the process. For a more complete evaluation of the portfolio value and the choice of the best specific action, management costs and times should be collected together with other data. This is crucial to determine the effective value of the recovery flow and, consequently, to identify the best collection process. At the same time, the data quality of recovery flows should improve, correctly assigning a date to each flow. As most collection process data is inputted manually, automatic data quality checks could be introduced in the process to improve data quality. Standard times and specific “answers” at

Issue n. 17 / 2019 51 Credit Risk

References [5] Gareth, J. and Witten, D. and Hastie, T. and Tibshirani, R. An Introduction to Statistical Learning with Applications. Springer, June 2013.

[1] Breiman, L. Out-of-Bag Estimation. [6] Grömping, U. Variable Importance Statistics Department, University Assessment in Regression: Linear of California, December 1996. Regression versus Random Forest. [2] Breiman, L. Random Forests. The American Statistician, Vol. 63, Statistics Department, University N. 4, pp. 308-31, November 2009. of California, January 2001. [7] Hastie, T. and Tibshirani, R. and [3] Ciavoliello, L. G. and Ciocchetta, Friedman, J. The Elements of F. and Conti, F. M. and Guida, I. Statistical Learning: Data Mining, and Rendina, A. and Santini, G. Inference, and Prediction. Springer, What’s the value of NPLs? Banca Second Edition, 2009. D’Italia, Eurosistema. Notes on Financial Stability and [8] Khieu, H. and Mullineaux, D. J. Supervision, April 2016. and Ha-Chin, Y. The Determinants [4] Fawcett, T. An introduction to ROC of Bank Loan Recovery Rates. Journal analysis. Pattern recognition letters, of Banking and Finance, Vol. 36, Vol. 27, Issue 8, pp. 861-874, June Issue 4, pp. 923-933, April 2012. 2006.

www.iasonltd.com 52 Argo Magazine

Appendix

A. Portfolio Comparison are quite similar. The analysis, in fact, shows a portfolio split (negative/positive events) consis- The exercise here considered focused on estimat- tent with the higher recovery rated of the “training models able to assess the absolute recovery ing set” (3% against 3.3% of real data, compared of dossiers, at least at aggregate level. Some- to the 2.4% against 2.9% of the whole portfolio); times, however, the analyst does not look for the same hit ratio of positive events in sample a precise assessment of recovery in absolute (20%) and a similar All-True ratio (95.0% versus terms, but he is more interested in a compared the value of 95.8% obtained from the portfolio evaluation, i.e. he may be interested to evaluate as a whole). the performance of a new portfolio relatively to A new multilinear model is estimated for one on which he already has some knowledge the Recovery Rate. In this case the geographic of. variable (which is not relevant) was not retained, To test this feature in the framework of this yielding a simpler model, which is summarized exercise, the given portfolio is divided into two in Table 23. sub portfolios: the “known portfolio” (training Even with a reduced set of recovery cases, set), what the analyst is acquainted with, and the model squeezes the forecast for the recovery the “new portfolio” (testing set), with different rate quite close to the average, resulting in a characteristics, summarized in Table 21. quite good fit (see Table 24). Thought the two portfolios include the same The model estimated presents a similar per- number of dossiers and present the same credit formance, even if developed on a halved num- exposure (11.45 mln e), the training set was ber of dossier, is then applied to the “new port- built to include a 27% larger number of recov- folio” to investigate if it can recognize its lower ered cases (350 versus 275), reflected in 60,000 collecting rate. The ability of the new model eof higher recoveries (32% higher than the test- to detect recovery cases on the “testing set” is ing one). reflected by Table 26. The expectations are that a model trained on Decreasing the non recovery component the “known portfolio” will be able to recognize from 3.0% to 2.7%, the model clearly reflects the lower recovery level of the “new portfolio”. the lower number of recovered dossier of the A new model is then estimated on the “new portfolio” and, though it makes some more “known portfolio” with the same approach seen mistakes than on the “known portfolio”, its over- in Chapter “Results and Conclusions”, yielding all performance, represented by an All-True of results similar to those detected for the whole 95.8%, is in line with those of the “known port- dataset. The model variables are summarized in folio” (actually, it is even slightly higher). Table 22, highlighting their relative importance. Although a more detailed enquiry of the The geographic variable (Macro Region) is kept recovered amounts may provide a stronger ev- into the model, even with a low relevance, as it idence of the ability of this approach to distin- represents the aggregating key. guish between different portfolio, the analysis On the basis of the same 10% cut-off adopted of the basic components of the model already previously, the probabilities of recovery at represent a clear evidence. dossier level are used to identify recovery events In conclusion, the framework presented is and the result is represented in Table 25. suitable to evaluate at some aggregate level, the Even if estimated on half the portfolio size, characteristics of a new portfolio with respect to the performance in identifying recovery events, the one it was trained on, known by the analyst.

TABLE 21: Training and testing portfolio proﬁles

Issue n. 17 / 2019 53 Credit Risk

TABLE 22: Logit model - Training set

TABLE 23: Recovery Rate - Training set

TABLE 24: Average Recovery Rate

TABLE 25: Extended Confusion Table - Training set

TABLE 26: Extended Confusion Table - Testing set

www.iasonltd.com 54 Argo Magazine

B. Model Test presents characteristics quite similar to those seen for the model estimated on the whole set, Since the whole dataset was used during in particular the bottom line (which, for the first the model estimate in Chapter “Development model, was equal to 98.1% and 20.0%, respec- Approach”, the out-of-sample test for the mod- tively, for OK on sample, while it was equal to elling approach presented requires to develop 95.8% for All-True), confirming similar perfor- a new model based on a subset of available mances for the new model. dossiers in order to leave the remaining part for Though it may not represent the best cut-off testing. The estimate follows the same proce- for the new model, this analysis is based on the dure seen for model comparison, but with the same 10% cut-off adopted previously to allow overall portfolio randomly divided into two un- an easier comparison with the overall model. even sets: the smaller one (one third of dossiers, At the cost of increasing the number of mis- randomly selected) for testing a model esti- takes, hence reaching a slightly lower All-True, mated on the other two thirds. A summary the RF approach on the same sample identifies of the profiles is given in Table 27. more recovered dossiers, as shown in Table 32. As shown in Table 28, due to random selec- With regards to estimates of the recov- tion of dossiers, the recovery characteristics of ery rates, like the initial model, the new one the two subsets are equivalent. Considering re- squeezes the values sufficiently close to the aver- covered dossiers only, the ratio of 13 to 23 is con- age resulting in a good fit, even with a reduced served both for exposure and recovery amounts, set of recovery cases (see Table 30). while the recovery rate is homogeneous for the To investigate its out-of-sample performance, two subsets. the new model is then tested on the reduced A new model is estimated on the training test set, including 213 recovery cases only. Table set with the same approach used in Chapter 33 below confirms that the performance level “Development Approach”, and the outcome is represented by All-True is maintained also out- summarized in Table 29. of-sample (95.4% to 95.9%). This is a subset of the variables included In addition, the portfolio generated by in the overall model of Chapter “Development model application is even closer to the real one Approach”. The geographical variables, whose in terms of recovered/non-recovered dossiers relevance was already very weak in the first (2.8% of recoveries to 3% in the out-of-sample model, became totally irrelevant with a smaller set, while 2.3% to 2.9% registered in-sample). development set. The in-sample performance is In conclusion, the modelling approach pre- condensed in Table 31. sented is suitable to classify NPL both in-sample Even if estimated on a smaller dataset, the and out-of sample, i.e. of new portfolios of the Confusion table obtained from the new model same type.

TABLE 27: Development and test dataset

TABLE 28: Exposure and Recovery ratios of train sets

Issue n. 17 / 2019 55 Credit Risk

TABLE 29: Testing model

TABLE 30: Average Recovery Rate

TABLE 31: Logit Extended Confusion Table - In-sample testing

TABLE 32: RF Extended Confusion Table - In-sample testing

TABLE 33: Logit Extended Confusion Table - Out-of-sample testing

www.iasonltd.com 56 FinTech

FinTechs and Challenger Banks: Old Business, Brand New Approach FinTech

About the Author

Antonio Menegon: Manager and Senior Risk Quant With six years of experience in Risk Man- agement and Consulting industries, he is currently leading the team of Business An- alysts and Financial Engineers at one big pan- European bank. Graduated in Mathematics from Università degli Studi di Padova, he has been continuously interested in new quant topics, focusing in the last years on Machine Learning and its application in ﬁnance.

This article was written in collaboration with Ilaria Biondo, who at the time was working for Iason Consulting.

www.iasonltd.com 58 Argo Magazine

FinTechs and Challenger Banks: Old Business, Brand New Approach aaaa

Ilaria Biondo Antonio Menegon

inTechs are increasingly challenging the status quo of the financial system, either providing brand new services or revisiting the actual players’ offerings. The rise of these companies is fueled by big financing F from PE, VC and crowdfunding, and by an extremely favourable ground from both a regulatory and a clientele point of view. Given this general scenario, the paper focuses on the major new fintech players in the European banking landscape, analysing their business model and the possible implications for the traditional commercial banks.

ith this paper the authors aim at pro- cally enabled financial innovation that results viding a general overview of the in new business models, applications, processes W new banking paradigma - the chal- or products with an associated material effect lenger or neo bank - that is stemming, in a on financial markets, institutions and the pro- nutshell, from FinTech development, new regu- vision of financial services. This means that latory frameworks and change in the customer through the application of technology, fintech behaviour. firms (hereafter simply FinTechs) can help to In the dissertation, the analysis will focus provide better, cheaper and faster services than primarily on the eurozone, where the traditional traditional commercial banks. commercial bank business model is compared to In the recent years, FinTechs saw their popu- the one adopted by new banking players which larity skyrocket year after year, mainly due to are rising. Among the latter, the authors will their ability to create easy and clear solutions to compare the approach of the biggest neo-banks, the major issues and inefficiencies customers are the English Revolut and Monzo and the German facing with traditional banking system. New N26. such companies are growing all over the world, The paper concludes with a comparison be- including emerging countries where there is a tween pros and cons of the two canvas, high- high percentage of unbanked people. FinTechs lighting possible issues and opportunities for can do so, even if not only, focusing on young both traditional and challenger banks. people and leveraging on technologies almost everyone owns like the smartphones, key channels to access fintech financial services provi- FinTech at a Glance sion. In 2018 global investments in FinTechs hit In this Section, the authors will present an more than $100 billion and 2,000 deals, sector overview of the FinTech sector, introducing the reports say. Some of global fintech start-ups major players on the market and where they are have reached huge valuations, exceeding $1 bil- concentrated. lion, fueling the search for firms to invest in that are more likely to have an impact on financial Overview of the FinTech Landscape industry, the so–called Unicorns. Looking at the top 100 FinTechs in the world, Citing the Financial Stability Board, in this pa- almost half is established in the Asia Pacific per we will refer to FintTech as a technologi- area,

Issue n. 17 / 2019 59 FinTech

FIGURE 17: Number of major FinTech companies; source: H2 Ventures [2]

FIGURE 18: Breakdown of the aforementioned 100 FinTechs by sector; source: H2 Ventures [2]

with China as leader in the market, while are using the latest financial technologies to im- 22 are American companies and 36 operates in prove the traditionally outdated and not always EMEA area (cfr. [2]) - see Figure 17. that transparent lending process, giving cus- Many of these companies are established in tomers the possibility to obtain customized ex- the payments sector (see Figure 18), focusing periences based on their own needs. Exploiting, on sending and saving money services or easy usually, artificial intelligence-related approaches, payments and transfers in any currencies. In such companies can gather and accurately pro- this regards, China is by far the largest market cess large amount of data, allowing them to with Ant Financial (part of Alibaba group) and grant clients with credit in as short as just few Tencent. Ant Financial’s Alipay and Tencent’s minutes. WeChat Pay have surpassed 500 and 900 million Innovative and data driven approaches like monthly active users respectively, or 36% and these helped, for example, Ant Financial to 65% of the overall population, accounting to- serve off-line farmers who could not provide gether for 94% of the mobile payments market sufficient documentation to apply for regular in China (cfr. [1]). bank credit. Although many of the FinTechs are currently concentrated in payments, another important Moving to US, we can find players like growing sector, especially in the US and Asian Avant, OnDeck or PeerIQ, which also use tech- markets, is the credit provisioning. FinTechs nology to simplify the loan application process:

www.iasonltd.com 60 Argo Magazine

- Avant allows loan-seekers to select dif- European payment directive PSD2 (Re- ferent loan options (debt consolidation, vised Payment Service Directive). The home improvement, emergencies,. . . ) and directive focuses on innovation improve- obtain up to $35,000 deposited into their ment and internet payment safety. In par- bank account even in just one day; ticular, customers must be able to use services (e.g. paying bills, making cash - OnDeck offers personalized loans to small transfers, ...) from different providers, as and mid-sized businesses, identifying the long as their money is deposited safely sector in which they operate and the pur- within their bank accounts. All banks pose of the loan in order to create a cus- are, then, required to create a system of tomized payment structure that best fits open APIs (Application Programming In- each situation; terfaces) that can ensure easy access to customers’ accounts to authorized third-party - PeerIQ allows an easier, transparent and providers, allowing them to offer different responsible decision-making process for financial services using the banks’ data. loan originators, asset managers and un- derwriters; The directive is basically aimed at levelling the playground and unifying the fi- Other minor sectors in which some players nancial market across Europe, at least re- in the fintech industry are specializing in are garding the payment and money transfer related to insurance, investments, savings and sectors, giving other parties legal and tech- brokerage. An example is Yu’ ebao, a Chinese nical tools to build their applications and online money market fund that was established services on top of banks’ data. All the to allow customers to invest small cash amounts providers with legal frameworks will be sitting in their Alipay accounts and that has be- allowed to operate not only in their home come, now, the largest money market fund in country but also in other countries of the the world. EU: a start-up in Italy, for example, may access the financial data in France as sim- FinTechs and Europe ple as the same data in Italy. In short, banks are required to be more Historically, with respect to other countries such open and focused on interaction with pay- as US or Asia, Europe has always been a step ment service providers (PSPs). backwards in the technological and digital innovations. However, in the recent years, the old To sum up, the directive is being a catalyst country has been an even more fertile ground for fintech innovation since it has been for FinTech development, as proved by the big implemented with two main purposes: excitement of venture capital investors about - Improving unification and competi- the European FinTechs; for example, according tion, thus lowering the cost of finan- to data from Dealroom: cial services and actually providing - Oaknorth raised e 783m; benefits to customers; - Creating open access to banks’ finan- - N26 raised e 612m; cial data in order to foster fintech development. - Atom bank’s raised e 486m. 2. The large volume of venture capital Most European FinTechs operate in the pay- investments, proven by the fact that in ment sector and market leaders like Revolut, 2018 investments in FinTechs companies Monzo, N26 or Transferwise have been included have increased up to $34.2 billion. in 2019 in the global Unicorn list, proving even more the dynamic nature of investments in this segment in the EU. In our view, distinctive reasons for active development of FinTechs in Europe can be traced to:

1. The favourable and stable regulatory environment, deﬁned primarily by the

Issue n. 17 / 2019 61 FinTech

FIGURE 19: Number of customers for the big 3 challenger banks in Europe from their inception; source: companies websites.

A New Business Model considered to add value to a long-term relationship with clients, they may be only available Following what has been briefly introduced, this for limited hours and sometimes crowded, mak- paragraph aims at presenting the differences ing it time-consuming to conduct even basic between the business model of conventional transactions. and challenger banks, describing their customer base, channels, main revenues streams and costs, Revenue streams. In a nutshell, the source highlighting how neo-banks are redefining the of revenues of a traditional commercial bank way of banking. can be summarised in three major streams: Among challenger banks, we chose to an- 1. Interests on loans, the principal income of alyze the big 3 in the European FinTech land- these type of banks whose primary role is scape: Revolut, Monzo and N26. the financial intermediation, expressed by the credit provisioning to private clients Traditional Commercial Banks and corporates;

A traditional bank operates mainly with physi- 2. Fees applied to services, mainly for the ac- cal branches and leverages on stratified legacy counts management (credit cards, money systems to provide the customers with all the transfer and exchange, ...); services in their offering. Such a bank has at its core the personal relationship with the clients, 3. Distribution channel commissions, basi- usually built on reciprocal trust. cally fees applied by the bank in the place- ment of, for example, investment and in- Customer Base. Traditional banks own a surance products. large pool of customers built on several years of operations in the financial industry. The clients Costs. The main costs a traditional bank has are actually well diversified in nature, with the to bear can be summarized in: banks providing offerings and services to pri- vates (all the different age segments) and corpo- 1. Interests payed on deposits; rates. 2. Costs for manage all the different branches; Channels. Traditional banks are organized with physical branches where customers can 3. Maintenance of the huge legacy systems; open a bank account and access banking ser- 4. Costs due to regulatory compliance. vices. The branches represent the primary contact point with customers: although they are

www.iasonltd.com 62 Argo Magazine

Challenger Banks Revolut Overview First to be analysed is Revolut, which debuted in 2015 and, since then, has attracted more than 8 New competitors for traditional banks are rep- million members. The firm has got its ECB bank- resented by the so-called challenger banks or neo- ing license in December 2018, which allowed the banks, small to medium sized enterprises with a English Fintech to ensure more protection on digital-first approach in service delivery which customers’ deposits (under the European De- are operating in the financial and banking indus- posit Insurance Scheme) and offer all services try. Neo-banks provide a wide stock of banking typically provided by conventional banks. This features, mainly in services related to sending has represented its first step towards globaliza- and saving money. At the core, these players tion. allow to open either free basic or premium ac- The company has been a step ahead of com- counts, which provide customers with the ac- petitors thanks to its highly flexible contracts cess to multiple services at even more conve- and free international cash transfers. It of- nient fares; just to name a few: fers free basic and different premium accounts, where the latter provide customers with addi- - Free payments in any currency; tional services and benefits (such as travel insur- - Cryptocurrencies purchases; ance, device insurance, metal card,..), depending on the kind of premium contract subscribed. - Insurance cover; Customer Base. Almost half of the com- - Free withdrawals worldwide. pany’s user base is between 25 and 35: target customers are young people, early adopters and Challenger banks operate with a new and in- businesses (70,000 companies signed up for Rev- novative business model, which takes into con- olut for Business, mainly freelancers, start-ups sideration trends in today’s society and aims and SMEs). at fulfilling customers’ current needs: transparency, flexibility, convenience. Digital is the Channels. The strategy adopted to acquire, only predominant channel for engaging clients, retain and continuously develop its customers without any physical branch on the territory. is coherent with its fintech nature. Channels are The key is designing all the services and the for the most part digital, including social media, user experience to be as straightforward as pos- tube ads and referral campaign activated by the sible: it all starts with downloading the app. community. Challenger banks do start as common Fin- Revolut’s core value, in fact, is represented Techs, and only in a second moment receive by its online community, highly focused on the the banking license through an application pro- products offered by the bank and their virality. cess which starts with the national competent authorities. The assessment is then, for the eu- Revenue streams. We can summarize Rev- rozone case, conducted jointly with the ECB olut’s revenues in two major streams: which is the competent authority for making 1. The premium account subscriptions; the final decision to grant, extend or withdraw a banking license. An authorization decision 2. Ratchet fees applied by the firm. Rev- must be taken within 12 months and the assess- olut is free/low-fee, however customers ment is, generally, focused on increased liquid- are time-limited and rate-limited (e.g. fee ity requirements recognizing the volatility of charged only on weekends or limited free FinTechs activity. monthly withdrawal up to $400 for pre- In Europe we can find major players like Rev- miums). Moreover, the company raises olut, Monzo and N26, which have managed to money from services provision such as challenge the traditional banking system by ex- P2P loans or cryptocurrency trading and ploiting technological innovations and achieve from business accounts with added op- a national and global outlook for much less cost tions for enterprises. and time compared to what took traditional banks. Costs. The main costs Revolut has to face to The next three paragraphs focus on the run the business are related to technology devel- aforementioned players, as key examples in the opment, regulations, compliance and licensing, changing banking system in Europe. marketing and to community management.

Issue n. 17 / 2019 63 FinTech

To sum up, what Revolut offers to its cus- Revenue streams. Monzo’s revenues can tomers is a redeﬁned way of banking with: be traced to 3 different streams:

- No/low banking fee; 1. Premium accounts subscriptions, providing benefits, such as travel insurance, ac- - Frictionless/feeless FX currency transfers cording to the account chosen; across multiple global currencies; 2. Provision of overdraft protection to cus- - Branchless banking, strong digital app tomers for a small monthly fee; and easy money transfers across different mediums. 3. Interest-bearing savings/ISA accounts in partnership with OakNorth Bank, on Monzo which Monzo earns a certain percentage One of the fastest-growing players in the euro- depending on the product. zone is the other UK-based firm, Monzo, started As plan for the future, Monzo is working on in 2015 and now counting almost 3.5 million launching also business accounts, which com- of customers. The FinTech got its banking li- petitors (i.e. Revolut and N26) already offer. cense in 2017, becoming a fully-fledged bank and ensuring more protection for customers’ Costs. Similar to the other challenger banks, money thanks to Financial Services Compensa- Monzo’s main costs are represented by technol- tion Scheme (FSCS). ogy development, marketing, customer support One of the flagship services the company and new partner integrations. Also, they have to bet on to stand from competition is the set of face expenses related to regulations, compliance features to easily track expenses and manage and licensing. money: users can break down the outflows into different categories, being able to control how To conclude, the company has created a new they spend their balances to the last coin. More- marketplace-model of digital banking which: over, similar to other players, Monzo offers fee- free international transactions and ATM with- - Helps users save money, manage budgets, drawals up to $200 per month and provides make easy transfers; customized overdrafts to eligible clients. - Offers savings/ISA accounts, travel insurance and other services thanks to partner- Customer Base. One-third of Monzo’s cus- ships; tomer base is represented by users aged between 25-31 years. The bank targets millennials, - Provides a top-notch customer support. mainly savers, providing services for budgeting, splitting bills and managing money which are N26 popular across generations. Moreover, Monzo’s customers are not simply “consuming” bank’s N26 is the other, not UK based, major player services every month; in fact, thousands of them, in Europe, having recently reached 3.5 million the so-called “prosumers”, have also invested users from foundation in 2013. The German in the company via equity crowdfunding cam- firm received its license in 2016, key turning paigns over the last 3 years. point in the corporate history: the banking license helped the firm grow its customer base Channels. Consistent with its digital na- across Europe by 1 million clients in two years, ture, the bank is all app-based, with no physical enabling the company to operate in 24 different branches available. Like the competitor Revo- markets. lut, Monzo has built a strong community forum N26 is popular thanks to its wide range of used to answer FAQs and basic questions. More- banking services, easy to access from every loca- over, the bank offers to customers channels into tion. One of the key features is the possibility of other providers for more financial services or fee-free transactions when using the card from insurance products. The company has indeed any part of the world: the app won’t charge created strategic partnerships in other niches conversion fees when using different currencies, (i.e. savings accounts with OakNorth bank) and the perfect solution for frequent international in new markets (i.e. US with Sutton Bank). travellers.

www.iasonltd.com 64 Argo Magazine

FIGURE 20: Comparison between the general business model of traditional commercial banks and challengers.

Customer Base. The company has a strong Revenue streams. Just as Revolut, the focus on youth-driven lifestyle brand for bank- company offers free basic accounts and two- ing. In fact, even in this case, targeted customer tiered subscription options which include bene- segments involve mostly young Europeans, but fits such as free withdrawals abroad, discounts also freelancers and self-employed. Many of with certain partners or a sleek card. Also in these customers have been acquired thanks to this case the main revenue streams are: strategic partnerships (which facilitated N26 ex- - Premium accounts subscriptions; pansion), other targeted, as in the case of the American market, thanks to a consistent mar- - Usage fees, applied depending on the ac- keting campaign. count chosen; - Business accounts, which create additional Channels. The relationship with customers revenues, based on value-added services, is all digital-based: there are no physical such as travel insurance or purchase insur- branches and customer service channels are on- ance. line platform or phone services; social media, of course, represent the most exploited resource to Costs. Main costs for the company are rep- reach customers and maintaining the relation- resented by technology development, marketing ship created. strategies, customer service and, of course, reg- Opposed to competitors though, N26 seems ulations, compliance and licensing. to be more commercial-partner focused and less community-driven: its approach to marketing To sum up, what N26 offers to its customers and customer support is more traditional. In is banking for native mobile generation: order to increase scalability and efficiency of - No/low banking fee; the business, the firm has partnered with com- - Lifestyle brand, offering benefits like part- panies like Mastercard, which provides all the nership discounts to youth brands; issuing and processing solutions on the payment side, or Wirecard, which provides banking - Business accounts for freelancers and self- back-end. employed.

Issue n. 17 / 2019 65 FinTech

Business Models Comparison reasons. At least at the moment, conventional banks are indeed assumed to be more stable and Challenger banks are revolutionizing the way solid and therefore more trustworthy; in addi- people interact with their own money. They tion, they are recognized brands with greater have improved the concept of customer service, knowledge of regulation with respect to chal- raised convenience in cash transfers and loan lenger banks, new on this field. application and automatized several financial In general, it is possible to summarize as fol- services like simple bill payments. lows some negative aspects proper of challenger In this section the authors summarize the banks that traditional banks do not present, main benefits and drawbacks seen in the new from customers’ and company’s point of views: proposed fintech business model. Main Benefits of Challenger Banks. It is - Customer’s point of view: clear that challenger banks provide a lot of benefits to customers who are looking for a more * Possible Limited Services: unlike convenient, accessible and easy-to-use solution. fully fledged traditional banks, chal- Such benefits, in the authors view, can be sum- lenger ones could be more special- marized as follows: ized, thus offering only limited services such as savings and budgeting - Improved Customer Service and Experi- or mortgages; ence: better customer support on phone * Less Solid and New on Financial and through multiple access points like Markets: this makes neo-banks less social media, usually providing feedback trustworthy than conventional ones in a sensible shorter time; (i.e. no reliable P&L history is available); - Easy Access to Banking Services: challenger banks are primarily app-based * Data Exposure: most challenger banking setups, meaning that it is always banks may be more prone to share 19 possible to access the desired services customer data with third parties . from any part of the world, through any - Company’s point of view: device and with disposal for more time during the day/ week; * Cyber Threats: online ventures are more exposed to cyber risk; - Additional Services provision and customized financial planning advice: most * Limited Capital: traditional banks neo-banks provide new kind of services can access more capital than neo that conventional banks do not offer (e.g. ones, at least at inception; robo-advisors); * Lower Loyalty among Customers: simple access to different banking - No/Low Banking Fee: all fees and providers makes it easy for cus- charges applied are aimed to be afford- tomers to switch whenever it be- able and highly transparent; comes more convenient; - Less Bureaucracy: easy account opening * New Regulation: regulators may and fast processing of banking services; take action and introduce some con- straints which could undermine the - Advanced Options to Control Credit fintech innovative business model Cards: new features allow customers to, and, therefore, compromise some dis- for example, freeze, unfreeze or geo-lock tinctive competitive advantages chal- their own card. lenger banks have.

Main Drawbacks of Challenger Banks. All the advantages just presented come of course with some drawbacks. Traditional banks have on their side a wide pool of loyal clients who, especially the older ones, are not willing to switch, maybe mainly for safety or legacy 19This risk is, at least, in part mitigated thanks to the new GDPR (General Data Protection Regulation).

www.iasonltd.com 66 Argo Magazine

Conclusions more their growth and been able to reach easier the new generation. The FinTech revolution is challenging the tradi- - Regulation. In order to foster competi- tional banking model and can threat the exist- tion for the ultimate goal of lowering the ing players which so far maintained their domi- costs of credit, competent authorities al- nant position thanks to different factors, from lowed players outside the banking and a highly regulated environment to immature financial industry to approach the exist- technology. ing pool of clients, often not that satisfied with the current offering. Key Takeaways - Big VC/PE Investments. In an era where In the authors’ view, key factors that led new start-ups can be the new gold, VCs and banking players to rise and gain marketshare in PEs have invested even heavily in new such a short time are mainly: ideas. Capital can be then at disposal to more small companies that, also with -A dvanced and Accessible Technology. crowdfunding, can reach sufficient critical Increasingly powerful hardware led to the mass to operate and start growing. advent of BigData and AI, empowered then by the scalability of Cloud comput- Challenger banks are thriving in this envi- ing. It all paved the way for a radical ronment, offering old services in a restyled way change in the paradigm: a company has and adding on top of it new services, all at lower no more to heavily invest in IT systems costs. They can do that mostly for what we just to elaborate data and extract from them stated: lighter infrastructures and organization, insightful business ideas. new customers (both private and businesses) looking for a more agile way to connect with Such a change shed definitely the light banks and less employees per customer (see on one of the biggest issues traditional figure 21 for a quite nice example of it). banks have: huge and no more that competitive legacy systems. These carry with them a burden in terms of both money Final Remarks to be invested for periodic updates and Given all of that, are traditional commercial people to be budgeted in order to take banks doomed? Definitely hazardous and too care of running, updating and changing early to say so. the IT infrastructure (usually fragmented One of the best solutions for banks could in a myriad of smaller architectures, every be to cooperate with FinTechs, which may add one of them with its own language, logics, value to traditional banking system by provid- data and owners). ing flexibility, innovation, cost reduction, greater The possibility for almost everyone to ex- ability to use data and a better user experience. ploit scalable computing power and (usu- Banks could retain as their core business the ally) open source softwares has enabled banking license, the customer database and the start-ups to proliferate in the shadows of compliance activity, producing a limited range the big players, often not able to keep the of products and services while the others would pace of much smaller and agile compa- be offered through APIs by FinTechs. Examples nies. are already in place especially in the US were big players in the banking sector partnered with - Digital Channels and New Customer BigTechs: Goldman Sachs with Apple and Citi Behaviour. Nowadays, people is used to with Google, just to name two cases. a connected and social media-oriented life, In this scenario (the “Distributed Bank”, cfr. where the great part of the day gravitates [1]), traditional and challenger banks will not around connected devices. Easy and fast compete for the ownership of customers rela- consumption of data and information are tionship, but they will operate jointly in the what the target clients do expect, are com- delivery of services. fortable with and looks for. Given these On the other side, banks could take the lead premises, start-ups which leveraged on in new technologies: they could focus on tar- digital channels, such as for example Face- geting customer niches that look for innovative book and Instagram, have often boosted services and specific products; however, this

Issue n. 17 / 2019 67 FinTech

FIGURE 21: Ratio of customers to employees between Monzo and RBS, chosen as representative of challenger and traditional banks respectively in the UK banking system; source: https: // murdo. xyz/ monzo-difference.

may be difficult for banks with large legacy IT systems. Moreover, risks may include not being able to find suitable personnel or partners for a successful innovative project and being subject to changing customer tastes and needs. In the authors’ view, traditional banks still have a great competitive advantage: their customer pool, clients which are not yet ready to pull all their money out of the accounts they already have, and the regulatory supervisors which will not risk a brutal disruption. After all, no robust P&L history is available for the new players, which are still in their early ex- pansion phase, with all the risks it concerns. If existing players do modernize, acquire (even merge) and/ or exploit partnerships with the new “cool kids in town”, a new banking system could arise, hopefully profitable for banks, Fin- Techs and the final customers.

www.iasonltd.com 68 Argo Magazine

References Evolution, Beneﬁts, Risk and Oversights. Argo Magazine N.16, July 2019.

[2] H2 Ventures. FinTech100. 2019.

[1] Biondo I. and Menegon A. [3] Lumos Business. Strategy for the BigTech and New Banking Landscape - Digital Economy.

www.iasonltd.com 69 Market Risk

Security Market: an Overview of Repo and Security Lending Transactions Argo Magazine

About the Authors

Nicola Giancaspro: Business Analyst Experienced as a business analyst in an international banking context, he has been supporting the development of the counterparty risk chain by offering functional support, coming from the background of studies and work experience gained in over two years.

Francesco Zorzi: Quantitative Analyst As a quantitative analyst, he has been working on the supporting activity for a major Italian global bank, dealing with the counterparty credit risk framework. He has also supported new regulation implementation on the aggregation phase of the counterparty credit risk chain.

This article was written in collaboration with Giammarco Dalessandro, who at the time was working for Iason Consulting. www.iasonltd.com 71 Market Risk

Security Market: an Overview of Repo and Security Lending Transactions aaaa

Giammarco Dalessandro Nicola Giancaspro Francesco Zorzi

ith this work, we would like to provide an overview of the Repurchase Agreement (Repo) and the Security Lending markets. Often, it could be difficult to distinguish Repo and Security Lending: W analyzing both typologies of instruments we first look for any differences for what concerns their technical structures, economic profile, participants, legal arrangements. Second, we analyze how these kind of instruments behaved and which vulnerabilities appeared during the last financial and banking crisis. We end the article focusing on structured typologies of Repo, tri-party trade and the total return swap highlighting characteristics in common with classic Repo.

epo markets are an essential source of se- the financial market distress, focusing on the cured financing for banks and financial main vulnerabilities encountered by the partic- R institutions and are a key instrument for ipants; finally the third Chapter is focused on the implementation of monetary policy. A repo, structured agreement that seems to be hybrid or sale and repurchase agreement, is a sale of solutions to be put in place by the market agents. security joined with an agreement to repurchase In the end, we will summarize the differences. the same security at a specified price at the end of the contract. Otherwise, securities lending involves the owner of shares or bonds transferring Repo and Security Lending them temporarily to a borrower, in return, the borrower transfers other shares, bonds or cash, Repo Instruments to the lender as collateral and pays a borrowing The Repo market is one of the worldwide largest fee. Both securities lending and repo are classi- segments of the money market and it is impor- fied under the Securities Financing Transaction tant for many reasons: an economic agent can (SFT) typology. The two types of instruments fulfill its necessity of funding, liquidity manage- have many similarities and can often be used ment, and investment needs. In the last decade, as functional substitutes for each other. Even this market has been developing and covering if they both are secured/collateralized loans more asset classes such as corporate bonds, and their definitions could appear to be similar, sovereign bonds, emerging market bonds, eq- some differnces between the two deal typolo- uity, and equity baskets. Key features for their gies can be found. The aim of this paper is to success could be linked to the fact that Repos investigate the main differnces arising between have a simple structure and are flexible, so they these two deals in order to clarify how they have become very popular among many partici- are distinguished by the market agents. The pants (central banks, investment banks, institu- work is organized as follows: the first Chapter tional investors and fund managers). They have highlights the repo and security lending main also gained popularity since they push liquidity economic aspects, the market participants and into the inter-bank market and decrease costs the legal arrangements on which they rely; in for the capital issuer and, in particular, they al- the second Chapter we found some differences low the market makers to cover their positions analyzing how both products behaved during more efficiently.

www.iasonltd.com 72 Argo Magazine

FIGURE 22: Classic Repo transaction.[3]

FIGURE 23: Sell/Buy-Back transaction.[3]

Classic Repo Repo is a money market instrument and it provides various benefits: As a derivative contract, a repurchase agreement is the sale of a security with a commitment by • Market makers are generally able to fi- the seller to buy the same security back from the nance their long positions as well as the purchaser at a specified price at a designated short one, both on bond and equity trade future date. For example, a dealer who owns a with a lower cost; 10-year U.S. Treasury note might agree to sell this security (the “seller”) to a mutual fund (the • During a bond issue, the REPOs are “buyer”) for cash today while simultaneously demonstrated to be a good way to bring agreeing to buy the same 10-year note back at liquidity to the overall operation; a predetermined price at a certain date in the • They are used by central banks for mon- future or, in some cases, the repurchase will be etary policy transmission through open on-demand. As a consequence, a repurchase market operations; agreement is a collateralized loan where the collateral is the security that is sold and subse- • Investors have one more investment op- quently repurchased. Repo contracts can also tion when allocating funds; be used to borrow securities. In this case, the • collateral provider earns a return by investing Institutional investors which may be the cash it receives from the cash investor at a forced to detain certain amount of security higher rate than that implied by the repo con- position for a long term, can use them in tract. Figure 22 is a schematic representation of repo transactions to get cash and use this a classic repo. liquidity for other purposes.

Issue n. 17 / 2019 73 Market Risk

Classic Repo Sell/Buy-Back “Sale” and repurchase Outright sale; forward buy-back Bid at repo rate: bid for stock, lend the cash Repo rate implicit in forward buy-back price Offer at repo rate: offer the stock, take the cash Sale and repurchase prices identical Forward buy-back price different Return to cash lender is repo interest on cash Return to cash lender is difference between sale price and forward buy-back price Bond coupon received during trade is returned to Coupon need not be returned to bond seller until seller termination Standard legal agreement (GMRA) No standard legal agreement (could be under GMRA) Initial margin may be taken Initial margin may be taken Variation margin may be called No variation margin, unless transacted under a legal agreement Speciﬁc repo dealing systems required May be transacted using existing bond and equity dealing systems

TABLE 34: Summary of features of Classic Repo and Sell/Buy-Back.[3]

Sell/Buy-Back vest cash. Typical cash investors are money market mutual funds and cash collateral reinvest- A different typology of Repo is represented by ment accounts managed for securities lenders the sell/buy-back deal which is another impor- and corporate treasuries, as well as financial tant money market instrument. It consists of institutions, such as banks, securities dealers, an outright sale of a bond on the value date equities, and derivatives exchanges. and an outright repurchase of the same bond The cash investors enter a repo trade to earn a re- on a forward date (i.e. sale of the bond at spot turn while having some protection, represented price and repurchase of the bond at the forward by the collateral, against losing their principal in price). Figure 23 represents a sell/buy-back cases of counterparty default. Otherwise, cash transaction. borrowers enter into repo contracts to finance The mentioned forward price will include the their securities positions or obtain leverage. interest on the repo, so the gain is realized Firms such as hedge funds typically engage a leveraging on the difference between the two securities dealer to access the repo market. Se- prices (spot price and forward price). In case of curities dealers provide collateralized financing sell/buy-back the repo rate is not explicit but to their clients and repledge securities collateral implied in the forward price and any coupon to obtain funding from cash investors. payments during the term are paid through in- Central banks are also an important customer corporation into the forward price, so the seller for repo business since they use them as a tool will receive them at the end. of monetary policy to control liquidity in the The sell/buy-back is used for the same reasons domestic money market. A central bank “repo” as a classic repo, but it was initially developed operation is actually a reverse repo, as it buys where no legal agreement existed to cover repo an eligible securities, typically domestic govern- transaction or where some counterparties were ment debt, and lends out cash to a list of eligible not ready to deal with classic repo. counterparties. The net effect is a short term in- In Table 34 features of both classic repo and jection of cash which mitigates shortages in the sell/buy-back transaction are presented. money market. Central banks have usually conducted temporary open market operations by Market Participants and Motivations entering into repo and reverse repo transactions The main participants to the REPO market are with primary dealers. agents who are willing to lend cash and agents who are seeking funding. Moreover, the market Legal Arrangments is not moved just by the need to lend/borrow cash but also by the need of those looking to A repo deal is legally structured as a syn- earn extra yield by lending securities and those chronous agreement between two parties en- looking to borrow specific securities. Typically, gaged in a sale of a security on an initial date, the former group includes buy-and-hold asset with a repurchase of the securities by the initial managers (such as pension, mutual, and in- seller at a later date. The model legal agreement surance funds) while the latter includes short- designed for parties dealing with Repos is the sellers (such as hedge funds). Cash lenders (or GMRA. cash investors) use repo as a way to securely in-

www.iasonltd.com 74 Argo Magazine

FIGURE 24: Security Lending transaction.

GMRA stands for Global Master Repurchase from the normal operation of most of the insol- Agreement and it is published by the Interna- vency laws. This has some consequences: when tional Capital Market Association (ICMA). As a counterparty defaults and gets broken, the an organization representing the cross border counterparty holding the security or the cash repo market in Europe, it provides agents le- lender has the right to liquidate the held secu- gal documentation used across all international rity. markets. The document is a pre-printed master agree- Security Lending ment and it essentially consists of three macro- Basic Concepts areas: • Standard provisions which are the gen- Securities finance involves secured borrowing eral conditions and set the standard REPO and lending transactions that are motivated for transaction; various reasons such as obtaining securities for settlement, financing inventory positions, and • Annex I: which represents the specific producing incremental returns by lending se- choices that should be made by the parties curities. In the U.S. equity markets, securities to put in place the agreement; lending has gained primarily importance after • Supplemental terms which are part of the the interdiction on “naked” short selling, which Annex I and where supplementary terms is a short sale by an institution that does not and conditions are provided and recorded hold the security and therefore cannot complete whenever the parties want to customise the delivery. This has created a “market” for the agreement and reflect specific needs; securities lending, which allows an institution that wants to sell a security and take a short • Annex II: where the commercial terms of position, to borrow it. This is how security lend- each transaction are recorded in confirma- ing is used in the equity market, while, in the tion. fixed-income markets, it is used not only for The GMRA is designed for short-term repos short selling but also for other borrowing deals of simple high-quality fixed-income securities such as security-for-security arrangements. that take the form of repurchase trade between An institution may also want to borrow a secu- principals under the law of England and Wales. rity to hedge risk through the use of derivatives To apply the GMRA to repos of equities, re- or to avoid “failing” on a delivery. Securities pos by or with an agent, or repos in the form lending is a transaction whereby a market par- of buy/sell-backs, it is necessary to amend the ticipant borrows an asset for a certain period. master agreement. This can be done by signing The aim of the security borrower is to avoid a the standard Equity, Agency, and Buy/Sell-Back delivery failure or to cover (or open) a short Annexes, respectively. Annex I can be easily position with respect to that asset. At the same modified with special supplementary terms or time, the goal of the securities lender is to earn a conditions when it is needed to adapt the agree- commission, increasing the return on his securi- ment to other jurisdictions. ties portfolio. In Figure 24 a scheme could help Despite a repo is structured as a sale, its eco- to have general view of the whole transaction nomic effect is similar to a secured loan. How- Securities lending today has a major role in the ever, unlike a secured loan, a repo transac- efficient functioning of the securities markets tion provides significant protections to creditors worldwide. Such transactions, in fact, are col-

Issue n. 17 / 2019 75 Market Risk

lateralized and the “rental fee” charged, along • Custodian banks: they are able to mobi- with all other aspects of the transaction, is dealt lize large pools of securities that are avail- under the terms agreed between the parties. able for lending. Involving them could This has the following consequences: have many advantages such as existing banking relationship with their customers, • Security and collateral ownership is their investment in technology and global passed between the parties; coverage of markets, the ability to pool assets from many smaller underlying funds, • The economic benefits such as dividends insulating borrowers from the administra- or coupon got from the security have to tive inconvenience of dealing with many be passed to the original owner of such small funds and providing borrowers with security, even if possession is on the bor- protection from recalls. They are also able rower; to provide indemnities and manage cash • Lender loses the right to vote whenever collateral efficiently. the security is equity type. • Third party agents: they make it possible These deals can be traded through two kinds of to separate the administration of securi- facilities: the “automatic securities lending facil- ties lending from the provision of basic ity” provided by a clearing house to its members custody services, and a number of spe- and the “special securities lending facility” pro- cialized third-party agency lenders have vided by a DMO to its PDs. established themselves as an alternative to An automatic securities lending facility is an the custodian banks. arrangement between a clearing house and its members, the goal of which is to avoid delivery • Principal intermediaries: they are compre- failures. Meanwhile, the purpose of a special hensive of dealers and brokers, insurance securities lending facility is to help PDs comply companies, hedge and pension funds. In with their market-making obligation. contrast to the agent intermediaries, prin- Most securities lending is done against cash col- cipal intermediaries can assume principal lateral. Typically, the lender of security will pay risk, offer credit intermediation, and take an interest rate to the borrower for the cash col- positions in the securities that they borrow. lateral. If security is rare or limited, then the A further role of the intermediaries is to interest rate paid by the lender will typically be take on liquidity risk. lower. In addition to the return potentially gen- • Beneficial owners/lenders: pension funds, erated through the lending transaction, lenders insurance, and assurance companies, mu- of securities seek to earn an additional return tual funds, unit trusts, and endowments. by investing the cash collateral. As a low-margin business, the lending portfolio needs to be of sufficient size to Market Participants make a securities lending program eco- The securities lending market involves various nomic. A relatively static portfolio with types of specialized intermediaries which take low securities turnover is more attractive principal and/or agency roles. They can be to securities borrowers because it mini- classified as follows: mizes recalls of loaned securities. Security may be recalled when its beneficial owner • Agent intermediaries together with the would like to sell it or exercise its voting owner will be dividing revenues from se- rights. curities lending at market rates. The split will be defined by many factors includ- • Security borrowers: the primary securities ing the service level and provision by the borrowers are securities dealers, who bor- agent of any risk mitigation, such as an row for their market-making activities or indemnity. on behalf of their clients. Dealers, which often act as market-makers, borrow secu- • Asset managers: especially in Europe, rities to settle buy orders from customers. where custodian banks were slower to A lack of securities to borrow may result take up the opportunity to lend, they are in less liquid markets with wider bid-ask a quick solution to reinvest the cash ob- spreads. The execution of many trad- tained from a financing operation. ing strategies relies on the ability of the

www.iasonltd.com 76 Argo Magazine

trader to borrow securities. For example, Under such dealings, the collateral is segregated traders often borrow securities to estab- into an account with a third-party custodian in lish a short position in one security that the name of the borrower, then making it the has been taken to hedge a long position in subject of the security interest in favor of the another security. lender, separating it from the lender’s assets and guarding it against the risk of non-return Legal Arrangments on the bankruptcy of the lender. The collateral is in the form of securities, finan- Securities lending is a global business and has cial instruments or cash. As either the value legal arrangement that can differ between ju- of the collateral, on one hand, or the value of risdictions. According to international practice, the loaned securities, on the other, shifts and the lender of the loaned securities gives the le- transfers are made in and out of the Secured gal title of the loaned securities to the borrower Account. If the collateral balance surpasses the for the duration of the loan. The lender retrieves demanded collateral value on a business day, the title at the end of the operation when the the lender notifies the custodian to release to the securities are rendered. Although the lender borrower from the Secured Account collateral gives up juridical ownership, the economic ben- which has a market value as close as possible to efit of any corporate operations, such as a stock the amount of the excess. If the collateral bal- split or income payments connected with the ance drops below the required collateral value loaned security, are retained by the lender: any on a business day, the borrower transfers into income or dividends are passed through from the Secured Account further collateral which the securities borrower to the lender. However, has a market value as close as possible to the in the case of equity securities, the lender loses amount of the deficit. any voting rights associated with the security In addition to these requirements, the documen- during the term of the loan. tation also covers: In Europe the master agreement used for such • Security Agreement: a security agreement a transaction is the GMSLA reproduced by between the lender as a secured party and the International Securities Lending Associa- the borrower as a security provider, which tion (ISLA) which publishes market standard creates the security interest over the collat- documents for security lending deals. eral in the Secured Account; For each loan of securities, the borrower is required to provide collateral transferring assets • Control Agreement: tri-party control to the lender. Commonly the borrower is ex- agreement between the lender, the bor- pected to over-collateralize the lender to guaran- rower and the custodian, whose main tee that the lender is fully preserved upon the function is to regulate access to the Se- risk of the borrower defaulting on the return cured Account to ensure that the lender of the loaned securities. As an outcome, if an has sufficient control over the Secured Ac- event of default happens to either borrower or count to reinforce its security interest. lender and the transactions under the GMSLA are closed out under its close-out netting mech- Borrowers and lenders also need to enter into a anism, the net termination payment would typi- security agreement to create protection over the cally be owing by the lender to the borrower and collateral in favor of the lender. This agreement should be equal to the return of the excess col- can follow distinct jurisdictions, these are: lateral. If the borrower is a financial institution, • English security agreement; its claim on the lender for the return of excess title-transfer collateral is a risk-weighted asset • Luxembourg law agreement: it depends for regulatory capital schemes, which demands on where the account is opened (JP Mor- allocation of capital and therefore has an impact gan or Clearstream); on the borrower’s balance sheet. However, if • Belgian law agreement for account opened the collateral is given transferring security, the at Euroclear. borrower retains a property interest in the collateral assets and is not exposed to the same risk These kind of agreement then differ on secu- of non-return by the lender, therefore its return rity interest, secured obligations, enforcement does not carry such a risk weighting. events, remedies, financial collateral arrange- The security collateral agreement is, therefore, a ments. charming prospect, in particular for borrowers.

Issue n. 17 / 2019 77 Market Risk

REPO and SEC Lending the first time since the beginning of the biannual survey in 2001, there were more banks with de- during the Crisis clining than with expanding repo books. The market also saw a severe shortening of maturi- Vulnerabilities of Repo Market ties and a total reduction of the volume of repo This Section briefly highlights developments in transactions by 26%, which was the biggest drop repo markets during the crisis. Repo markets recorded since the survey began. play a key role in facilitating the flow of cash However, with the crisis unfolding, uncertainty and securities around the financial system. They emerged on the value of the collateral provided offer a low-risk and liquid investment for cash, in these deals. Besides, there was rising uncer- as well as the efficient management of liquid- tainty about the liquidity of markets on which ity and collateral by financial and non-financial collateral such as asset-backed securities could firms. A well-functioning repo market also sup- be sold if the counterparty defaulted on the repo ports liquidity and price discovery in cash mar- loan. In general, counterparty risk grew. kets, helping to improve the efficient allocation The spread of the Euribor to the Overnight In- of capital and to reduce the funding costs of dexed Swap, being an indicator of counterparty firms in the real economy. However, excessive risk on inter-bank markets, increased consider- use of repos can facilitate bubbles, excessive ably in mid-2007 when the subprime market leverage and encourage reliance on short-term collapsed: the widening spread was connected funding. to the rescue of some banks such as Bear Stearns. Even if the European REPO Market was orig- inally less interconnected with the sub-prime The Role of Global Banks During the Spread related asset class and the most popular under- of the Crisis lying security were sovereign securities, doubt Vulnerabilities in the repo and short-term large- on the market liquidity and counterparty sol- scale funding markets have been mentioned by vency spread as well. It was standard practice policymakers and regulators as a latent source in both Europe and in the US for collateral to of systemic stress. Weaknesses were especially be “rehypothecated”, suggesting that the col- emphasized during the 2007-09 financial crisis lateral received in a repo agreement could be when over-dependence on wholesale funding used by the lender in another repo transaction. had a role in the collapse of Bear Stearns Com- The dominance of bilateral repo agreements is panies, Lehman Brothers and Britain’s Northern another feature of the European repo market Rock. The Financial Stability Oversight Council which might have played a part in spreading (FSOC) has highlighted various risks associated skepticism among market participants. More with the repo market and suggested some ac- than 50% of repo trades in Europe are carried tions to enhance the structure of the tri-party out on a bilateral basis. repo market and limit latent spillovers from repo-related asset fire-sales. Repo and Leverage Financing In [6] the author has highlighted the role of globally active banks in transmitting the crisis, Following [9] the extreme use of repos in the which began with the collapse of the US sub- creation of leverage and financing long-term as- prime housing market and spread to the global sets with short-term funding was one factor that financial system. The disruptions in the US contributed to the Great Financial Crisis. The financial system triggered a worldwide finan- crisis demonstrated that this type of funding cial crisis that continued to increase in terms of can be extremely volatile and can quickly disap- spreading for more than two years following the pear in times of market or idiosyncratic stress. collapse of the subprime market in mid-2007. The strongest adverse repercussion on intra- Such negative outcomes would not have been bank lending provoked by the repo funding possible without further transmission channels shock can be noticed after the Bear Stearns res- besides the direct exposure of banks around the cue, the event which raised even bigger con- world to the US subprime market. cerns about the solvency of potential counter- The inter-bank refinancing system uses securi- parties in the inter-bank lending market. An tized assets as collateral in sale and repurchase example of a different kind of run seems to (repo) transactions, and it was believed to be have happened when institutions that relied on less risky than unsecured funding. The Interna- the repo market for its funding have been driven tional Capital Market Association states that, for into bankruptcy whenever its creditors declined

www.iasonltd.com 78 Argo Magazine to extend repo financing. This had happened current approach to daily unwinds could con- to some major banks in the tri-party repo mar- tribute to behavior by cash investors or clearing ket during the crisis, as lenders reacted to the banks that leads to an unexpected loss of financ- perceived creditworthiness of the counterparty ing to one or more dealers. For example, assume quickly selling assets to meet higher collateral that cash investors start to mature concerns that requirements. a clearing bank may refuse to unwind the repos Following [8] it seems in a bilateral market, of a certain dealer. Because that would force the which is mostly the case of Europe, haircuts dealer into default, cash investors would be un- increased rapidly and reached high levels. For willing to fund the dealer the day before. The this reason, the authors assert a run occurred failure of the dealer into obtaining financing since less cash could be raised from the borrow- from cash investors would moreover force the ers. Otherwise, following [4] the tri-party repo dealer into default. Similarly, a clearing bank market behaved differently during the crisis and would oppose itself to unwind the repos of a this may be due to the fact haircut have been dealer that is not likely to get financing from stable in such a period, suggesting this market cash investors at the end of the day. This would could be less risky. Anyway until now, this dif- therefore also be a self-fulfilling expectation. ference has not been well explained. It may be related in part to the lower average degree of Collateral Liquidation Frictions leverage of cash lenders in the tri-party market relative to those of the bilateral market. Another weakness of the tri-party repo market is the current lack of effective and clear plans to support the systematic liquidation of a default- Market Infrastructural Weaknesses ing dealer’s collateral. This has been moderated The second source of vulnerability is found by the central banks stepping in and providing in the weaknesses in the repo market’s in- a huge amount of liquidity for dealers to pro- stitutional infrastructure. A stable and well- vide them with a second source of financing functioning tri-party repo market is critical to and to guarantee clearing banks and lenders. the health and stability of the financial markets and economy. Some guide work has been made Vulnerabilities in Security Lending Market by the US Tri-party Repo Infrastructure Reform The security lending market starts with the ne- Task Force which can be summarised in the cessity to dodge the interdiction of the so-called following points: “naked selling” which is a short sale by an in- • Reduce discretionary intraday credit ex- stitution that does not hold the security and tended by tri-party clearing banks; therefore cannot complete delivery. The ban on naked short selling produces a role for securities • Promote improvements in market partic- lending, which allows an institution that wants ipant’s liquidity and credit risk manage- to sell a security short to borrow it. ment practices; Liquidity Transformation Risks • Reduce the risk of destabilising fire sales in the event of default. As in the repo markets, features of the securities lending market acted differently during the re- Nonetheless, some evidence shows how the cent crisis. A wide deleveraging took place, cre- clearing banks usually unwind all repos every ating liquidity stress and, in some cases, losses morning by 8:30 AM, whether they are matur- for securities lenders as they were obliged to ing that day or not. As a result, clearing banks return the cash collateral to the borrowers of extend intraday credit to dealers for the total the securities. The liquidity pressure and the value of the tri-party repo market, which was losses were typically aligned with the degrees about $2.8 trillion at the peak of the market and of credit risk and liquidity transformation as- was $1.6 trillion in May 2011. The large credit sociated with the investment of cash collateral. exposures that clearing banks have to dealers, There is a link between repo and security mar- although secured by the dealers’ assets, is a sys- ket: the securities lending market for govern- temic risk, given the size and centrality of the ment bonds lets borrowers upgrade collateral by clearing banks in the financial system. Whether exchanging risky assets such as corporate bonds, induced by concerns over a dealer’s solvency equities, or other products for high-quality liq- or over the credit quality of the collateral, the uid government bonds (i.e., safe assets) that

Issue n. 17 / 2019 79 Market Risk

are adequate for the repo market and central or return an equivalent amount of cash to their counterparties. Borrowed bonds are reused into account. Indemnification can also be applied to the market, serving as a source of financing, the reinvestment of cash collateral. for example, in the repo market, and the (cash) This fact lead to highlight another dangerous collateral can be reinvested in the money mar- factor of such a market: indemnification set- ket. The crisis surrounding AIG offers an exam- tlements expose lending agents to contingent ple. Like many other large insurance companies, claims. It is crucial that the party providing the AIG was involved in securities lending opera- indemnification is well-capitalized and holds ex- tivity. Before the financial crisis, its loans were cess liquidity to meet the possible demands of mostly open and its supply of cash collateral indemnification in the event of a counterparty was invested in particularly long-term and illiq- default. It is also equally important that the uid assets. This meant that AIG was working party providing the backstop to the transaction with a considerable liquidity transformation, uses a diversified list of counterparties to mini- which can result in liquidity stress. This invest- mize exposure to any one borrower. ment strategy allowed high returns before the Post-crisis regulatory changes, in particular crisis; however, it contributed to AIG’s liquidity with Basel III capital standards, require bank- squeeze during the crisis. The firm underwent affiliated lending agents to incorporate capital something similar to a run as borrowers of its charges, liquidity requirements and counter- securities tried to get them back as part of the party concentration limits to account for risks general market deleveraging taking place. The inherent in securities lending transactions. Since need to convert some illiquid assets into liquid these changes can make securities lending pro- ones to accommodate this return of securities grams more costly to run, one potential outcome contributed to a sizable share of AIG’s losses. could be a migration of securities lending ac- As a consequence of the financial distress and tivities away from banks to entities unaffiliated following the analysis of [1], it seems during with banks. crises lenders prefer to hold high-quality government bonds unless the lending fee is exces- sively high and borrowers are less likely to use cash, pledging non-cash collateral to borrow Structured Repo, Tri-Party high-quality government bonds of core coun- Repo and the Total Return tries during stressed periods. Swap Indemnification and Operational Risks The Repo market, as most of the financial in- To manage such a risk, a standard market prac- dustry, has been influenced by the financial en- tice that has been developed over the past sev- gineering and its structured products. In this eral decades and became a legal requirement or Chapter we briefly review some examples of a necessity for many investors, see agent lenders non-classic Repo and introduce the topic of the providing securities replacement guarantees or tri-party repo. We then look at the total re- indemnification for borrower default. Usually, turn swap (TRS), a credit derivative instrument this kind of protection occurs when collateral for which emerged many connections between is not able to cover the charge for buying again REPO products and security lending deals. at the market price the security. Many agents offer their customers more levels of indemnifi- Structured Repo cation options for their lending and collateral reinvestment programs. Indemnification is a Cross-Currency Repo vital part of the agent lending business and it is This trade is a Repo transaction where the collat- a key consideration for many investors engaged eral is denominated in a different currency with in securities lending. Indemnification means respect to the one in which the cash part is. For that, in the event of a counterparty default, the example, a repo may involve borrowing EUR agent would first use the available collateral against U.K. government bonds (bonds denomi- (typically collateralized from 102-105%) to re- nated in GBP). Since this transaction is exposed purchase the client’s securities or to return an to forex risk, it needs to be evaluated frequently equivalent amount of cash to their account. If to take into account foreign exchange rate fluctu- the collateral were insufficient to make the in- ations. Amending based on currency volatility vestor whole, then the agent lender would use helps ensure adequate collateralization of both its capital to repurchase the client’s securities

www.iasonltd.com 80 Argo Magazine

FIGURE 25: Tri-party Repo structure.[3] cash and securities. Banks and other financial its maturity and then continue funding beyond institutions use this type of repo to earn poten- the initial period. The two parties can keep get- tial funding gains in a currency different from ting the contract renewed, though it still can the one of the collateral. be ended with a specific time notice (such as 30-day notice), the so-called notice period. Callable Repo Basket Repo A callable repo is a transaction in which the repo rate is fixed and the lender of cash has the Basket repo is a repo of a portfolio of bonds. right to end the agreement before the maturity Banks and other financial institutions often repo date, or call partially back the cash. A callable out entire portfolios of bonds with a repo mar- repo has an embedded interest rate call option ket maker. It is operationally more convenient that can be exercised by the lender when nec- due to the fact that it is considered as one repo essary. This option will be exercised if interest trade. The mechanics of a basket repo are iden- rates rise during the life of the repo: in this case, tical to the ones of a classic repo. the lender can call back the debt to reinvest it at the higher market rates. Due to the presence Tri-Party Repo of the embedded call option and then to the payment of a premium, the lender will receive Tri-party repo is a repurchase agreement trans- a fixed rate lower than a standard repo rate. action for which post-trade operations is handled by a third-party agent. Post-trade opera- Open Repo tions consist of (but are not limited to): A repurchase agreement which has no specific • Collateral selection; repurchase date. The term of an open repo is • Payments and deliveries; not predefined, and therefore it has no end date. Open repo products allow banks and other fi- • Custody of collateral securities; nancial institutions to buy securities without having to stick to a specified repurchase date. • Collateral management. Either party to the agreement could terminate A tri-party agent can be a custodian bank, an in- the agreement at any time after the contract ternational central securities depository (ICSD) date. Due to the uncertainty associated with or a national central securities depository (CSD). open repos, their interest rates are commonly In Europe, the principal tri-party agents are higher than ordinary repos due to the presence Clearstream Bank Luxembourg, Euroclear Bank, of risk premium arising from not knowing the Bank of New York Mellon, JP Morgan and SIS. end date of the trade. In the US, there is now only one agent (Bank of New York Mellon) since JP Morgan largely gave Evergreen Repo up its tri-party agent role in 2018. In Figure 25 A type of repo which extends for longer periods is represented the basic structure of a tri-party than usual. It comes with an option to extend repo trade.

Issue n. 17 / 2019 81 Market Risk

FIGURE 26: Total Return Swap structure.[3]

Tri-Party Agent • Substitution at the request of the seller. Tri-party agent is just an intermediary, so the Another very important task of a tri-party agent use of a tri-party repo does not affect the risk is to conduct collateral optimization, which relationship between the two counterparties. If means the reassessment of collateral composi- one of the parties defaults, the impact falls on tion if there is a better combination of securities the other party. This reflects the fact that the that can be posted as collateral given changes two counterparties of a tri-party repo still have in the seller’s holdings, since the start of a tri- to sign bilateral legal agreements such as the party repo and execution of the substitutions ICMA Global Master Repurchase Agreement necessary to achieve the better combination. (GMRA). On the other hand, since collateral is typically Moreover, the tri-party agent doesn’t provide selected automatically by the tri-party agent, tri- a trading venue where the parties can nego- party repo cannot be used for borrowing and tiate and execute transactions (although some lending specific securities. This is reflected in tri-party agents are linked to trading platforms). the large average deal size of tri-party repo and Instead, once a transaction has been agreed di- collateralization by multiple securities. rectly between the parties, they independently notify the tri-party agent, who matches the in- Cost Effective with Drawbacks structions and processes the trade. An important feature is that, typically, the agent will au- Thanks to the aforementioned tri-party collat- tomatically select, from the securities account eral automated operations, advantages from the of the seller counterparty, collateral that satis- agent’s economies of scale and because settle- fies pre-agreed credit and liquidity criteria, con- ment is across the books of the agent, the op- centration limits and any transaction decisions erational costs of a tri-party repo are less than agreed between the buyer and the seller. those of a repurchase agreement handled in- Afterward, the tri-party agent handles the reval- house and settled across a securities settlement uation of the collateral, variation margining, in- system, which charges a fee for each securities come payments on the collateral, and substitu- transferred. This makes it economic to collat- tion related activities which includes: eralize a tri-party repo with multiple securities. Moreover, tri-party agents also can efficiently • Substitution of any collateral which ceases manage baskets of collateral denominated in to conform to the quality criteria of the several currencies. The ability to collateralize buyer; with multiple securities facilitates larger deal sizes, taking down the costs. • Substitution to prevent an income pay- On the other hand, the lower cost of a tri-party ment triggering a tax event; repo is an incentive to use non-government se- • Substitution of a security used as collateral curities as collateral. These less liquid securities which has been sold; trade in smaller amounts than government se-

www.iasonltd.com 82 Argo Magazine curities, this lesser liquidity can make bilateral is more common. Since there was concern transfers across securities settlement systems about the intra-day systemic risk in the a lot more expensive. Consequently, repos of U.S. tri-party market, there are some on- equity, corporate bonds, MBS, ABS, and other going reforms which will bring it closer to structured securities are concentrated in the tri- the European tri-party model. party repos market. • The US tri-party market is dominated by Players two types of investors, money market mutual funds and securities lending agents Tri-party market is the preferred repo market reinvesting cash collateral, who handle al- segment for many customers since the offload most two-thirds of tri-party market. These of collateral management to a tri-party agent investors are required or prefer to reinvest allows these firms to avoid the cost of setting most of their cash in repo and they tend to up and running their own collateral manage- use tri-party repo due to operational ad- ment department. The customer list includes vantages. The problem is that cash collat- central banks, some of whom allow the use of eral taken in the securities lending market tri-party agents by the counterparties in their is an open-ended liability (as the securi- monetary policy operations and others who use ties loans can typically be recalled at any tri-party services when conducting investment time) but most tri-party repos are collater- operations. alised by medium or long-term securities. In case of a default on a repo, investors European vs. US Tri-Party Markets would have to take the securities onto their balance sheets. Given that they cannot or Below we listed the most relevant differences may not wish to hold such longer-term between European and US tri-party markets. collateral securities, they would be forced or might feel impelled to immediately sell • Tri-party agents dominate the settlement those securities. If the default was by a of US repo, they handle almost two-thirds large borrower, sufficient collateral might of the outstanding volume of the US mar- be sold to trigger a fire sale, fuelling a self- ket, compared to about 10% in the Euro- reinforcing cycle of disposal and price col- pean market. lapse. The European tri-party repo market • European tri-party repo is normally used doesn’t suffer from such a concentration to manage non-government bonds and eq- of the investor base. uity (please note that the proportion of government bonds has more than doubled Total Return Swap since the Great Financial Crisis), while US tri-party is focused on Treasury and A total return swap (TRS) is an agreement be- Agency debt. tween two parties that exchanges the total return from a financial asset between them. It is • In most European tri-party systems, there one of the principal instruments used by banks has always been true term repo, while and other financial instruments to manage their term repos in US tri-party systems were credit risk exposure. As such, it is a credit traditionally unwound each morning and derivative product, but it has economic simi- re-arranged in the afternoon. This was larities with respect to repo. intended to give sellers the daily opportunity to withdraw and replace collat- Basic Concepts eral securities and adjust for price fluctuations, instead of going through the opera- One definition of a TRS is given in [7] which tionally more intensive direct substitution states: and variation margining with the other party. The downside of this procedure is “A total return swap is a swap agree- that it required the tri-party agents to fi- ment in which the total return of a bank nance the sellers for most of the day and loan(s) or credit-sensitive security(s) is this generates a systemic intra-day credit exchanged for some other cash flow, usu- exposure. In Europe, instead, the use of di- ally tied to LIBOR or some other loan(s) rect substitution and variation margining or credit-sensitive security(s)”

Issue n. 17 / 2019 83 Market Risk

The underlying asset could be either physically instrument, a party can reduce exposure to an sold to the other counterparty, with a transac- asset without having to sell it. tion agreed by the two parties or there could Moreover, since the maturity of the swap be no physical exchange of the underlying secu- doesn’t have to match the one of the reference rity. Furthermore, the TRS trade has no need to asset, the swap receiver may gain from the pos- match the maturity of the underlying security. itive funding that derives from being able to In a TRS the total return from the underly- roll over short-term funding of a longer-term ing asset is paid over the counterparty in re- asset. The total return payer, on the other hand, turn for a fixed or floating cash flow. This benefits from the protection against market and feature makes the TRS slightly different from credit risk for a specified period of time, with- other credit derivatives, as the payments be- out having to sell the asset. On maturity of the tween counterparties to a TRS are connected to swap, the total return payer could decide if it changes in market value of the underlying asset, continues to own the asset or sell it in the open as well as change resulting from the occurrence market. Thus the instrument may be considered of a credit event. a synthetic repo. In figure 26 is illustrated a generic TRS, bank A TRS traded as a synthetic repo is usually used A has to pay the “total return” on a specified to temporary remove a specific assets from the reference asset, while receiving a Libor-based balance sheet. This may be desirable for several return from bank B. The underlying asset can reasons: if the institution is due to be analysed be a bank loan such as a corporate loan or a by credit rating agencies or if the annual exter- sovereign or corporate bond. The total return nal audit is not too far away. payments from bank A include the interest pay- Another reason a bank wants to temporarily ments on the underlying loan as well as any remove lower-credit-quality assets from its bal- appreciation in the market value of the asset. ance sheet is if it is in danger of breaching capi- Bank B will pay the Libor-based return and it tal limits between quarterly return periods. In will also pay any price difference if there is a this case, as the return period approaches, lower- depreciation in the price of the asset. quality assets may be removed from the balance The economic effect of this structure is as if bank sheet by signing a TRS which will mature after B owns the underlying asset, so TRS could be the return period has passed. considered synthetic loans or securities. An important feature is that bank A will usually hold the underlying asset on its balance sheet so, if Synthetic Repo via Total Return Swap the TRS requires the physical exchange of the Synthetic repos, traded to fund a portfolio or underlying security, this means that the security borrow stocks to cover short positions, are com- will be removed from bank B balance sheet for mon in the market. In this case the repo is in the term of the TRS. the form of a Total Return Swap (TRS), which is The total return on the underlying asset is the a credit derivative instrument. However, when interest payments and any change in the market used for the aforementioned purposes, it is iden- value if there is an increase in price. The price tical in economic terms to a classic repo. TRS increase could be cash settled, or alternatively is similar to a synthetic repo contract in the there could be physical delivery of the reference sense that the economic effect is the same but asset at the maturity of the swap, in return for they are considered different instruments. The a payment of the initial asset value by the total biggest difference between the two products is return “receiver”. the following: TRS takes the underlying off the If the issuer of the reference asset defaults, the balance sheet, while the tax and accounting au- swap may be ended immediately, with a net thorities treat repo as if the underlying remains present value payment which sign depends on the balance sheet. Also, the agreement un- on the fact that the value has increased or de- der which a classic repo and a TRS are traded creased over time. is different: TRS is conducted under the ISDA standard legal agreement, while classic repo is Uses of Total Return Swap conducted under the GMRA standard repo le- There are several reasons why banks and fi- gal agreement. nancial institutions may want to enter into TRS In the next page we present an example from agreements, the main one is to reduce credit [3] regarding the use of TRS to fund a portfolio risk. In fact, using TRS as a credit derivative of bonds, as a substitute for a repo trade.

www.iasonltd.com 84 Argo Magazine

Example: Synthetic Repo via Total Return Swap

Consider a financial institution such as a regulated broker-dealer that has a portfolio of assets on its balance sheet that it needs to obtain funding for. These asset are investment-grade rated structured finance bonds such as credit card ABS, residential MBS and CDO notes, and investment-grade rated convertible bonds. In the repo market, it is able to fund these at Libor plus 16 bps. That is, it can repo the bonds out to a bank counterparty, and will pay Libor plus 16 bps on the funds it receives. Assume that for operation reasons the bank can no longer funds these assets using repo. Instead it can fund them using a basket TRS contract. Under this contract, the portfolio of assets is swapped out to the TRS counterparty, and cash received from the counterparty. The assets are therefore sold off the balance sheet to the counterparty, an investment bank. The investment bank need to fund this itself - it may have a line of credit from a parent bank or it may swap the bonds out into the market. The funding rate it charges the broker-dealer will depend to a large extent on what rate the bank can fund the assets itself. Assume that the TRS rate charged is Libor plus 22 bps - the higher rate reflects the lower liquidity in the basket TRS market for non-vanilla bonds. At the start of the trade, the portfolio consists of five EUR-denominated convertible bonds. The broker-dealer enters into a three-month TRS with the investment bank counterparty, with a one-week interest rate reset. This means at each one-week interval, the basket is revalued. The difference in value from the last valuation is paid (if higher) or received (if lower) by the investment bank to the broker-dealer; in return the broke-dealer also pays one-week interest on the fund it received at the start of the trade. In it can be broken at one-week intervals and bonds in the reference basket can be returned, added to or substituted. The terms of the trade are shown below:

• Trade date: 24 March 2004;

• Value date: 26 March 2004;

• Maturity date: 28 June 2004;

• Rate reset: 31 March 2004;

• Interest rate: 2.2970% (this is one-week EUR Libor ﬁx of 2.077% plus 16 bps).

The combined market value of the entire portfolio is taken to be EUR 102,477,023.48. At the start of the trade, the five bonds are swapped out to the investment bank, who pays the portfolio value for them. On the first reset date, the portfolio is revalued and the following calculation confirmed:

• Old portfolio value: EUR 102,477,023.48;

• Interest rate: 2.2970%;

• Interest payable by broker-dealer: EUR 45,770.22;

• New portfolio value: EUR 107,532,194.64;

• Portfolio performance: +5,055,171.166;

• Net payment: broker-dealer receives EUR 5,009,400.94.

There has been no change in the prices of the ﬁve convertible bonds, but the broker-dealer has added an ABS security to the portfolio. In addition, there has been one week’s accrued interest on the original portfolio. This make up the new portfolio value. The rate is reset for value 2 April 2004 for the period to 9 April 2004. The new rate is 22 bps over the one-week EUR Libor ﬁx on 31 March 2004, an all-in rate of 2.252880%. This interest rate is payable on the new “loan” amount of EUR 107,532,194.64. This trade has the same goals and produced the same economic effect as a classic repo transaction.

Issue n. 17 / 2019 85 Market Risk

Conclusions the International Securities Lending Asso- ciation (ISLA). Moreover, in the security In this paper we have explored the Repo and lending deals the lender of the securities Security Lending markets and we tried to under- gives the legal title of the securities to the stand the basic differences in the two financial borrower for the duration of the loan. derivatives deals even if they seem to be similar • Participants: in this case emerges that at a first glance. In the first Chapter we mainly participants are very similar between the found that they differ in the following aspects: deals with some differences in the role • Economic: an essential difference among they play. Repo and securities lending is that the In the second Chapter we have focused on Repo market mostly uses bonds and other the main vulnerabilities both transactions have fixed-income instruments as collateral, shown during the financial crisis. From our while a significant segment of the secu- point view, this could be seen as a source of rities lending market uses equities, even if differences: while in the Repo market rehy- both of them are substantially collateral- pothecation, leverage financing, infrastructure ized loans. weaknesses and the friction on the collateral • Rationale: while the Repo users are moti- liquidation have emerged as risk factors, in vated by the need to borrow and lend cash, the security lending market weaknesses arisen security lending has gained primarily im- due to its characteristic indemnification process portance after the interdiction on “naked” which brings operational risk with it. short selling, even though are motivated Finally, in the third Section some structured and for various reasons such as obtaining secu- synthetic security transaction have been listed rities for settlement, financing inventory which usually show features belonging both to positions, and producing incremental re- the Repo and to security lending trades, as in turns by lending securities. the case of the Total Return Swap. In the specific case of a TRS, in fact, it is possible to reproduce • Legal: the GMRA for Repo transactions the economic effect of a classic Repo combining is designed for short-term repos of sim- it with the removal of the underlying asset from ple high-quality fixed-income securities the balance sheet. This is to highlight that even that take the form of repurchase trade be- if some differences have arisen between the tween agents, while the master agreement deals, the markets remain strictly connected to used for security lending, which usually each other and their products as well. has riskier underlying, is the GMSLA by

www.iasonltd.com 86 Argo Magazine

References [6] Federal Reserve Bank of New York Tri-Party Repo Infrastructure reform: a white paper. Federal Reserve Bank of New York, May 2010.

[1] Aggarwal, R. and Bai, J. and [7] Francis, J. and Frost, J. and Leaven, L. Sovereign Debt, Whittaker, J.G. Handbook of Credit Securities Lending and Financing Derivatives. McGraw-Hill, 1999. during the crisis. Centre for [8] Gorton, G. and Metrick, A. Financial Markets and Policy, 2015. Securitized Banking and the Run on [2] Choudhry, M. Structured Credit Repo. Journal of Financial Products: Credit Derivatives and Economics, June 2012. Synthetic Securitisation. John Wiley [9] Grill, M. and Jakovicka, J. and & Sons, 2010. Lambert, C. and Nicoloso, P. and [3] Choudhry, M. The Repo Handbook. Steininger, L. and Wedow, M. Butterworth-Heinemann, 2010. Recent developments in euro area repo markets, regulatory reforms and their [4] Copeland, A. and Martin, A. and impact on repo market functioning. Walker, M. Repo Runs: Evidence European Central Bank, Financial from the Tri-Party Repo Market. Stability review, November 2017. Federal Reserve Bank of New York, July 2011. [10] International Capital Market Association What is tri-party repo. [5] Düwel C. Repo funding and internal International Capital Market capital markets in the ﬁnancial crisis. Association, 2019. Deutsche Bundesbank, N. 16, 2013.

www.iasonltd.com 87 New Frontiers in Practical Risk Management

In the previous issue

Issue N. 16- 2019 aaaa RGO Issue N. 16- 2019 New FrontiersA in Practical Risk Management

FinTech

BigTech and New Banking Landscape – Evolution, Beneﬁts, Risks and Oversights

Stress Test

Are EBA Stress-Test Driving Banks into Different Business Models?

1 Market Risk

2019 FRTB Review: Main Interventions

Last issues are available at www.iasonltd.com/research/